maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/X86/vector-shuffle-variable-256.ll

782 lines
33 KiB
LLVM
Raw Normal View History

Revert "In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled." Reverting due to ARM MCJIT and MIPS LLD error. This reverts commit r289659. llvm-svn: 289667
2016-12-15 00:43:44 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; NOTE: Assertions have been autogenerated by update_llc_test_checks.py
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+avx | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2
;
; Unary shuffle indices from registers
;
define <4 x double> @var_shuffle_v4f64_v4f64_xxxx_i64(<4 x double> %x, i64 %i0, i64 %i1, i64 %i2, i64 %i3) nounwind {
; ALL-LABEL: var_shuffle_v4f64_v4f64_xxxx_i64:
; ALL: # BB#0:
; ALL-NEXT: pushq %rbp
; ALL-NEXT: movq %rsp, %rbp
; ALL-NEXT: andq $-32, %rsp
; ALL-NEXT: subq $64, %rsp
[SDAG] Revert r296476 (and r296486, r296668, r296690). This patch causes compile times for some patterns to explode. I have a (large, unreduced) test case that slows down by more than 20x and several test cases slow down by 2x. I'm sending some of the test cases directly to Nirav and following up with more details in the review log, but this should unblock anyone else hitting this. llvm-svn: 296862
2017-03-03 18:02:25 +08:00
; ALL-NEXT: andl $3, %esi
; ALL-NEXT: andl $3, %edi
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; ALL-NEXT: andl $3, %ecx
; ALL-NEXT: andl $3, %edx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: vmovaps %ymm0, (%rsp)
; ALL-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; ALL-NEXT: vmovhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
; ALL-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
; ALL-NEXT: vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0]
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; ALL-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: movq %rbp, %rsp
; ALL-NEXT: popq %rbp
; ALL-NEXT: retq
%x0 = extractelement <4 x double> %x, i64 %i0
%x1 = extractelement <4 x double> %x, i64 %i1
%x2 = extractelement <4 x double> %x, i64 %i2
%x3 = extractelement <4 x double> %x, i64 %i3
%r0 = insertelement <4 x double> undef, double %x0, i32 0
%r1 = insertelement <4 x double> %r0, double %x1, i32 1
%r2 = insertelement <4 x double> %r1, double %x2, i32 2
%r3 = insertelement <4 x double> %r2, double %x3, i32 3
ret <4 x double> %r3
}
define <4 x double> @var_shuffle_v4f64_v4f64_uxx0_i64(<4 x double> %x, i64 %i0, i64 %i1, i64 %i2, i64 %i3) nounwind {
; ALL-LABEL: var_shuffle_v4f64_v4f64_uxx0_i64:
; ALL: # BB#0:
; ALL-NEXT: pushq %rbp
; ALL-NEXT: movq %rsp, %rbp
; ALL-NEXT: andq $-32, %rsp
; ALL-NEXT: subq $64, %rsp
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; ALL-NEXT: andl $3, %edx
; ALL-NEXT: andl $3, %esi
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: vmovaps %ymm0, (%rsp)
[X86][SSE] Add support for combining VZEXT_MOVL target shuffles Includes adding more general support for the pattern: VZEXT_MOVL(VZEXT_LOAD(ptr)) -> VZEXT_LOAD(ptr) This has unearthed a couple of latent poor codegen issues (MINSS/MAXSS scalar load folding and MOVDDUP/BROADCAST load folding patterns), which will be fixed shortly. Its also reduced a couple of tests so that they no longer reach the instruction threshold necessary to be combined to PSHUFB (see PR26183). llvm-svn: 279646
2016-08-25 02:07:53 +08:00
; ALL-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; ALL-NEXT: vmovddup {{.*#+}} xmm0 = xmm0[0,0]
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
; ALL-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; ALL-NEXT: movq %rbp, %rsp
; ALL-NEXT: popq %rbp
; ALL-NEXT: retq
%x0 = extractelement <4 x double> %x, i64 %i0
%x1 = extractelement <4 x double> %x, i64 %i1
%x2 = extractelement <4 x double> %x, i64 %i2
%x3 = extractelement <4 x double> %x, i64 %i3
%r0 = insertelement <4 x double> undef, double undef, i32 0
%r1 = insertelement <4 x double> %r0, double %x1, i32 1
%r2 = insertelement <4 x double> %r1, double %x2, i32 2
%r3 = insertelement <4 x double> %r2, double 0.0, i32 3
ret <4 x double> %r3
}
define <4 x double> @var_shuffle_v4f64_v2f64_xxxx_i64(<2 x double> %x, i64 %i0, i64 %i1, i64 %i2, i64 %i3) nounwind {
; ALL-LABEL: var_shuffle_v4f64_v2f64_xxxx_i64:
; ALL: # BB#0:
[SDAG] Revert r296476 (and r296486, r296668, r296690). This patch causes compile times for some patterns to explode. I have a (large, unreduced) test case that slows down by more than 20x and several test cases slow down by 2x. I'm sending some of the test cases directly to Nirav and following up with more details in the review log, but this should unblock anyone else hitting this. llvm-svn: 296862
2017-03-03 18:02:25 +08:00
; ALL-NEXT: andl $1, %esi
; ALL-NEXT: andl $1, %edi
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; ALL-NEXT: andl $1, %ecx
; ALL-NEXT: andl $1, %edx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; ALL-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; ALL-NEXT: vmovhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
; ALL-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
; ALL-NEXT: vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0]
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; ALL-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: retq
%x0 = extractelement <2 x double> %x, i64 %i0
%x1 = extractelement <2 x double> %x, i64 %i1
%x2 = extractelement <2 x double> %x, i64 %i2
%x3 = extractelement <2 x double> %x, i64 %i3
%r0 = insertelement <4 x double> undef, double %x0, i32 0
%r1 = insertelement <4 x double> %r0, double %x1, i32 1
%r2 = insertelement <4 x double> %r1, double %x2, i32 2
%r3 = insertelement <4 x double> %r2, double %x3, i32 3
ret <4 x double> %r3
}
define <4 x i64> @var_shuffle_v4i64_v4i64_xxxx_i64(<4 x i64> %x, i64 %i0, i64 %i1, i64 %i2, i64 %i3) nounwind {
; AVX1-LABEL: var_shuffle_v4i64_v4i64_xxxx_i64:
; AVX1: # BB#0:
; AVX1-NEXT: pushq %rbp
; AVX1-NEXT: movq %rsp, %rbp
; AVX1-NEXT: andq $-32, %rsp
; AVX1-NEXT: subq $64, %rsp
[SDAG] Revert r296476 (and r296486, r296668, r296690). This patch causes compile times for some patterns to explode. I have a (large, unreduced) test case that slows down by more than 20x and several test cases slow down by 2x. I'm sending some of the test cases directly to Nirav and following up with more details in the review log, but this should unblock anyone else hitting this. llvm-svn: 296862
2017-03-03 18:02:25 +08:00
; AVX1-NEXT: andl $3, %edi
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: andl $3, %esi
; AVX1-NEXT: andl $3, %edx
; AVX1-NEXT: andl $3, %ecx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovaps %ymm0, (%rsp)
; AVX1-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX1-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX1-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: movq %rbp, %rsp
; AVX1-NEXT: popq %rbp
; AVX1-NEXT: retq
;
; AVX2-LABEL: var_shuffle_v4i64_v4i64_xxxx_i64:
; AVX2: # BB#0:
; AVX2-NEXT: pushq %rbp
; AVX2-NEXT: movq %rsp, %rbp
; AVX2-NEXT: andq $-32, %rsp
; AVX2-NEXT: subq $64, %rsp
[SDAG] Revert r296476 (and r296486, r296668, r296690). This patch causes compile times for some patterns to explode. I have a (large, unreduced) test case that slows down by more than 20x and several test cases slow down by 2x. I'm sending some of the test cases directly to Nirav and following up with more details in the review log, but this should unblock anyone else hitting this. llvm-svn: 296862
2017-03-03 18:02:25 +08:00
; AVX2-NEXT: andl $3, %edi
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: andl $3, %esi
; AVX2-NEXT: andl $3, %edx
; AVX2-NEXT: andl $3, %ecx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovaps %ymm0, (%rsp)
; AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: movq %rbp, %rsp
; AVX2-NEXT: popq %rbp
; AVX2-NEXT: retq
%x0 = extractelement <4 x i64> %x, i64 %i0
%x1 = extractelement <4 x i64> %x, i64 %i1
%x2 = extractelement <4 x i64> %x, i64 %i2
%x3 = extractelement <4 x i64> %x, i64 %i3
%r0 = insertelement <4 x i64> undef, i64 %x0, i32 0
%r1 = insertelement <4 x i64> %r0, i64 %x1, i32 1
%r2 = insertelement <4 x i64> %r1, i64 %x2, i32 2
%r3 = insertelement <4 x i64> %r2, i64 %x3, i32 3
ret <4 x i64> %r3
}
define <4 x i64> @var_shuffle_v4i64_v4i64_xx00_i64(<4 x i64> %x, i64 %i0, i64 %i1, i64 %i2, i64 %i3) nounwind {
; AVX1-LABEL: var_shuffle_v4i64_v4i64_xx00_i64:
; AVX1: # BB#0:
; AVX1-NEXT: pushq %rbp
; AVX1-NEXT: movq %rsp, %rbp
; AVX1-NEXT: andq $-32, %rsp
; AVX1-NEXT: subq $64, %rsp
[SDAG] Revert r296476 (and r296486, r296668, r296690). This patch causes compile times for some patterns to explode. I have a (large, unreduced) test case that slows down by more than 20x and several test cases slow down by 2x. I'm sending some of the test cases directly to Nirav and following up with more details in the review log, but this should unblock anyone else hitting this. llvm-svn: 296862
2017-03-03 18:02:25 +08:00
; AVX1-NEXT: andl $3, %edi
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: andl $3, %esi
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovaps %ymm0, (%rsp)
; AVX1-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX1-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: movq %rbp, %rsp
; AVX1-NEXT: popq %rbp
; AVX1-NEXT: retq
;
; AVX2-LABEL: var_shuffle_v4i64_v4i64_xx00_i64:
; AVX2: # BB#0:
; AVX2-NEXT: pushq %rbp
; AVX2-NEXT: movq %rsp, %rbp
; AVX2-NEXT: andq $-32, %rsp
; AVX2-NEXT: subq $64, %rsp
[SDAG] Revert r296476 (and r296486, r296668, r296690). This patch causes compile times for some patterns to explode. I have a (large, unreduced) test case that slows down by more than 20x and several test cases slow down by 2x. I'm sending some of the test cases directly to Nirav and following up with more details in the review log, but this should unblock anyone else hitting this. llvm-svn: 296862
2017-03-03 18:02:25 +08:00
; AVX2-NEXT: andl $3, %edi
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: andl $3, %esi
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovaps %ymm0, (%rsp)
; AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-NEXT: movq %rbp, %rsp
; AVX2-NEXT: popq %rbp
; AVX2-NEXT: retq
%x0 = extractelement <4 x i64> %x, i64 %i0
%x1 = extractelement <4 x i64> %x, i64 %i1
%x2 = extractelement <4 x i64> %x, i64 %i2
%x3 = extractelement <4 x i64> %x, i64 %i3
%r0 = insertelement <4 x i64> undef, i64 %x0, i32 0
%r1 = insertelement <4 x i64> %r0, i64 %x1, i32 1
%r2 = insertelement <4 x i64> %r1, i64 0, i32 2
%r3 = insertelement <4 x i64> %r2, i64 0, i32 3
ret <4 x i64> %r3
}
define <4 x i64> @var_shuffle_v4i64_v2i64_xxxx_i64(<2 x i64> %x, i64 %i0, i64 %i1, i64 %i2, i64 %i3) nounwind {
; AVX1-LABEL: var_shuffle_v4i64_v2i64_xxxx_i64:
; AVX1: # BB#0:
[SDAG] Revert r296476 (and r296486, r296668, r296690). This patch causes compile times for some patterns to explode. I have a (large, unreduced) test case that slows down by more than 20x and several test cases slow down by 2x. I'm sending some of the test cases directly to Nirav and following up with more details in the review log, but this should unblock anyone else hitting this. llvm-svn: 296862
2017-03-03 18:02:25 +08:00
; AVX1-NEXT: andl $1, %edi
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: andl $1, %esi
; AVX1-NEXT: andl $1, %edx
; AVX1-NEXT: andl $1, %ecx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; AVX1-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX1-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX1-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: retq
;
; AVX2-LABEL: var_shuffle_v4i64_v2i64_xxxx_i64:
; AVX2: # BB#0:
[SDAG] Revert r296476 (and r296486, r296668, r296690). This patch causes compile times for some patterns to explode. I have a (large, unreduced) test case that slows down by more than 20x and several test cases slow down by 2x. I'm sending some of the test cases directly to Nirav and following up with more details in the review log, but this should unblock anyone else hitting this. llvm-svn: 296862
2017-03-03 18:02:25 +08:00
; AVX2-NEXT: andl $1, %edi
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: andl $1, %esi
; AVX2-NEXT: andl $1, %edx
; AVX2-NEXT: andl $1, %ecx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: retq
%x0 = extractelement <2 x i64> %x, i64 %i0
%x1 = extractelement <2 x i64> %x, i64 %i1
%x2 = extractelement <2 x i64> %x, i64 %i2
%x3 = extractelement <2 x i64> %x, i64 %i3
%r0 = insertelement <4 x i64> undef, i64 %x0, i32 0
%r1 = insertelement <4 x i64> %r0, i64 %x1, i32 1
%r2 = insertelement <4 x i64> %r1, i64 %x2, i32 2
%r3 = insertelement <4 x i64> %r2, i64 %x3, i32 3
ret <4 x i64> %r3
}
define <8 x float> @var_shuffle_v8f32_v8f32_xxxxxxxx_i32(<8 x float> %x, i32 %i0, i32 %i1, i32 %i2, i32 %i3, i32 %i4, i32 %i5, i32 %i6, i32 %i7) nounwind {
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; ALL-LABEL: var_shuffle_v8f32_v8f32_xxxxxxxx_i32:
; ALL: # BB#0:
; ALL-NEXT: pushq %rbp
; ALL-NEXT: movq %rsp, %rbp
; ALL-NEXT: andq $-32, %rsp
; ALL-NEXT: subq $64, %rsp
; ALL-NEXT: # kill: %R9D<def> %R9D<kill> %R9<def>
; ALL-NEXT: # kill: %R8D<def> %R8D<kill> %R8<def>
; ALL-NEXT: # kill: %ECX<def> %ECX<kill> %RCX<def>
; ALL-NEXT: # kill: %EDX<def> %EDX<kill> %RDX<def>
; ALL-NEXT: # kill: %ESI<def> %ESI<kill> %RSI<def>
; ALL-NEXT: # kill: %EDI<def> %EDI<kill> %RDI<def>
; ALL-NEXT: andl $7, %edi
; ALL-NEXT: andl $7, %esi
; ALL-NEXT: andl $7, %edx
; ALL-NEXT: andl $7, %ecx
; ALL-NEXT: andl $7, %r8d
; ALL-NEXT: vmovaps %ymm0, (%rsp)
; ALL-NEXT: andl $7, %r9d
; ALL-NEXT: movl 16(%rbp), %r10d
; ALL-NEXT: andl $7, %r10d
; ALL-NEXT: movl 24(%rbp), %eax
; ALL-NEXT: andl $7, %eax
; ALL-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; ALL-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
; ALL-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
; ALL-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; ALL-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; ALL-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[2,3]
; ALL-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1],mem[0],xmm1[3]
; ALL-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1,2],mem[0]
; ALL-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; ALL-NEXT: movq %rbp, %rsp
; ALL-NEXT: popq %rbp
; ALL-NEXT: retq
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
%x0 = extractelement <8 x float> %x, i32 %i0
%x1 = extractelement <8 x float> %x, i32 %i1
%x2 = extractelement <8 x float> %x, i32 %i2
%x3 = extractelement <8 x float> %x, i32 %i3
%x4 = extractelement <8 x float> %x, i32 %i4
%x5 = extractelement <8 x float> %x, i32 %i5
%x6 = extractelement <8 x float> %x, i32 %i6
%x7 = extractelement <8 x float> %x, i32 %i7
%r0 = insertelement <8 x float> undef, float %x0, i32 0
%r1 = insertelement <8 x float> %r0, float %x1, i32 1
%r2 = insertelement <8 x float> %r1, float %x2, i32 2
%r3 = insertelement <8 x float> %r2, float %x3, i32 3
%r4 = insertelement <8 x float> %r3, float %x4, i32 4
%r5 = insertelement <8 x float> %r4, float %x5, i32 5
%r6 = insertelement <8 x float> %r5, float %x6, i32 6
%r7 = insertelement <8 x float> %r6, float %x7, i32 7
ret <8 x float> %r7
}
define <8 x float> @var_shuffle_v8f32_v4f32_xxxxxxxx_i32(<4 x float> %x, i32 %i0, i32 %i1, i32 %i2, i32 %i3, i32 %i4, i32 %i5, i32 %i6, i32 %i7) nounwind {
; ALL-LABEL: var_shuffle_v8f32_v4f32_xxxxxxxx_i32:
; ALL: # BB#0:
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; ALL-NEXT: # kill: %R9D<def> %R9D<kill> %R9<def>
; ALL-NEXT: # kill: %R8D<def> %R8D<kill> %R8<def>
; ALL-NEXT: # kill: %ECX<def> %ECX<kill> %RCX<def>
; ALL-NEXT: # kill: %EDX<def> %EDX<kill> %RDX<def>
; ALL-NEXT: # kill: %ESI<def> %ESI<kill> %RSI<def>
; ALL-NEXT: # kill: %EDI<def> %EDI<kill> %RDI<def>
; ALL-NEXT: andl $3, %edi
; ALL-NEXT: andl $3, %esi
; ALL-NEXT: andl $3, %edx
; ALL-NEXT: andl $3, %ecx
; ALL-NEXT: andl $3, %r8d
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; ALL-NEXT: andl $3, %r9d
; ALL-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; ALL-NEXT: andl $3, %r10d
; ALL-NEXT: movl {{[0-9]+}}(%rsp), %eax
; ALL-NEXT: andl $3, %eax
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; ALL-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
; ALL-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
; ALL-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; ALL-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[2,3]
; ALL-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1],mem[0],xmm1[3]
; ALL-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1,2],mem[0]
; ALL-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; ALL-NEXT: retq
%x0 = extractelement <4 x float> %x, i32 %i0
%x1 = extractelement <4 x float> %x, i32 %i1
%x2 = extractelement <4 x float> %x, i32 %i2
%x3 = extractelement <4 x float> %x, i32 %i3
%x4 = extractelement <4 x float> %x, i32 %i4
%x5 = extractelement <4 x float> %x, i32 %i5
%x6 = extractelement <4 x float> %x, i32 %i6
%x7 = extractelement <4 x float> %x, i32 %i7
%r0 = insertelement <8 x float> undef, float %x0, i32 0
%r1 = insertelement <8 x float> %r0, float %x1, i32 1
%r2 = insertelement <8 x float> %r1, float %x2, i32 2
%r3 = insertelement <8 x float> %r2, float %x3, i32 3
%r4 = insertelement <8 x float> %r3, float %x4, i32 4
%r5 = insertelement <8 x float> %r4, float %x5, i32 5
%r6 = insertelement <8 x float> %r5, float %x6, i32 6
%r7 = insertelement <8 x float> %r6, float %x7, i32 7
ret <8 x float> %r7
}
define <16 x i16> @var_shuffle_v16i16_v16i16_xxxxxxxxxxxxxxxx_i16(<16 x i16> %x, i32 %i0, i32 %i1, i32 %i2, i32 %i3, i32 %i4, i32 %i5, i32 %i6, i32 %i7, i32 %i8, i32 %i9, i32 %i10, i32 %i11, i32 %i12, i32 %i13, i32 %i14, i32 %i15) nounwind {
; AVX1-LABEL: var_shuffle_v16i16_v16i16_xxxxxxxxxxxxxxxx_i16:
; AVX1: # BB#0:
; AVX1-NEXT: pushq %rbp
; AVX1-NEXT: movq %rsp, %rbp
; AVX1-NEXT: andq $-32, %rsp
; AVX1-NEXT: subq $64, %rsp
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: # kill: %R9D<def> %R9D<kill> %R9<def>
; AVX1-NEXT: # kill: %R8D<def> %R8D<kill> %R8<def>
; AVX1-NEXT: # kill: %ECX<def> %ECX<kill> %RCX<def>
; AVX1-NEXT: # kill: %EDX<def> %EDX<kill> %RDX<def>
; AVX1-NEXT: # kill: %ESI<def> %ESI<kill> %RSI<def>
; AVX1-NEXT: # kill: %EDI<def> %EDI<kill> %RDI<def>
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovaps %ymm0, (%rsp)
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl 32(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: movzwl (%rsp,%rax,2), %eax
; AVX1-NEXT: vmovd %eax, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl 40(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $1, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl 48(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $2, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl 56(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $3, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl 64(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $4, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl 72(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $5, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl 80(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $6, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl 88(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $7, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $15, %edi
; AVX1-NEXT: movzwl (%rsp,%rdi,2), %eax
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovd %eax, %xmm1
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $15, %esi
; AVX1-NEXT: vpinsrw $1, (%rsp,%rsi,2), %xmm1, %xmm1
; AVX1-NEXT: andl $15, %edx
; AVX1-NEXT: vpinsrw $2, (%rsp,%rdx,2), %xmm1, %xmm1
; AVX1-NEXT: andl $15, %ecx
; AVX1-NEXT: vpinsrw $3, (%rsp,%rcx,2), %xmm1, %xmm1
; AVX1-NEXT: andl $15, %r8d
; AVX1-NEXT: vpinsrw $4, (%rsp,%r8,2), %xmm1, %xmm1
; AVX1-NEXT: andl $15, %r9d
; AVX1-NEXT: vpinsrw $5, (%rsp,%r9,2), %xmm1, %xmm1
; AVX1-NEXT: movl 16(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $6, (%rsp,%rax,2), %xmm1, %xmm1
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl 24(%rbp), %eax
; AVX1-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $7, (%rsp,%rax,2), %xmm1, %xmm1
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: movq %rbp, %rsp
; AVX1-NEXT: popq %rbp
; AVX1-NEXT: retq
;
; AVX2-LABEL: var_shuffle_v16i16_v16i16_xxxxxxxxxxxxxxxx_i16:
; AVX2: # BB#0:
; AVX2-NEXT: pushq %rbp
; AVX2-NEXT: movq %rsp, %rbp
; AVX2-NEXT: andq $-32, %rsp
; AVX2-NEXT: subq $64, %rsp
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: # kill: %R9D<def> %R9D<kill> %R9<def>
; AVX2-NEXT: # kill: %R8D<def> %R8D<kill> %R8<def>
; AVX2-NEXT: # kill: %ECX<def> %ECX<kill> %RCX<def>
; AVX2-NEXT: # kill: %EDX<def> %EDX<kill> %RDX<def>
; AVX2-NEXT: # kill: %ESI<def> %ESI<kill> %RSI<def>
; AVX2-NEXT: # kill: %EDI<def> %EDI<kill> %RDI<def>
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovaps %ymm0, (%rsp)
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl 32(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: movzwl (%rsp,%rax,2), %eax
; AVX2-NEXT: vmovd %eax, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl 40(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $1, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl 48(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $2, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl 56(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $3, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl 64(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $4, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl 72(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $5, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl 80(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $6, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl 88(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $7, (%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $15, %edi
; AVX2-NEXT: movzwl (%rsp,%rdi,2), %eax
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovd %eax, %xmm1
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $15, %esi
; AVX2-NEXT: vpinsrw $1, (%rsp,%rsi,2), %xmm1, %xmm1
; AVX2-NEXT: andl $15, %edx
; AVX2-NEXT: vpinsrw $2, (%rsp,%rdx,2), %xmm1, %xmm1
; AVX2-NEXT: andl $15, %ecx
; AVX2-NEXT: vpinsrw $3, (%rsp,%rcx,2), %xmm1, %xmm1
; AVX2-NEXT: andl $15, %r8d
; AVX2-NEXT: vpinsrw $4, (%rsp,%r8,2), %xmm1, %xmm1
; AVX2-NEXT: andl $15, %r9d
; AVX2-NEXT: vpinsrw $5, (%rsp,%r9,2), %xmm1, %xmm1
; AVX2-NEXT: movl 16(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $6, (%rsp,%rax,2), %xmm1, %xmm1
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl 24(%rbp), %eax
; AVX2-NEXT: andl $15, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $7, (%rsp,%rax,2), %xmm1, %xmm1
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX2-NEXT: movq %rbp, %rsp
; AVX2-NEXT: popq %rbp
; AVX2-NEXT: retq
%x0 = extractelement <16 x i16> %x, i32 %i0
%x1 = extractelement <16 x i16> %x, i32 %i1
%x2 = extractelement <16 x i16> %x, i32 %i2
%x3 = extractelement <16 x i16> %x, i32 %i3
%x4 = extractelement <16 x i16> %x, i32 %i4
%x5 = extractelement <16 x i16> %x, i32 %i5
%x6 = extractelement <16 x i16> %x, i32 %i6
%x7 = extractelement <16 x i16> %x, i32 %i7
%x8 = extractelement <16 x i16> %x, i32 %i8
%x9 = extractelement <16 x i16> %x, i32 %i9
%x10 = extractelement <16 x i16> %x, i32 %i10
%x11 = extractelement <16 x i16> %x, i32 %i11
%x12 = extractelement <16 x i16> %x, i32 %i12
%x13 = extractelement <16 x i16> %x, i32 %i13
%x14 = extractelement <16 x i16> %x, i32 %i14
%x15 = extractelement <16 x i16> %x, i32 %i15
%r0 = insertelement <16 x i16> undef, i16 %x0 , i32 0
%r1 = insertelement <16 x i16> %r0 , i16 %x1 , i32 1
%r2 = insertelement <16 x i16> %r1 , i16 %x2 , i32 2
%r3 = insertelement <16 x i16> %r2 , i16 %x3 , i32 3
%r4 = insertelement <16 x i16> %r3 , i16 %x4 , i32 4
%r5 = insertelement <16 x i16> %r4 , i16 %x5 , i32 5
%r6 = insertelement <16 x i16> %r5 , i16 %x6 , i32 6
%r7 = insertelement <16 x i16> %r6 , i16 %x7 , i32 7
%r8 = insertelement <16 x i16> %r7 , i16 %x8 , i32 8
%r9 = insertelement <16 x i16> %r8 , i16 %x9 , i32 9
%r10 = insertelement <16 x i16> %r9 , i16 %x10, i32 10
%r11 = insertelement <16 x i16> %r10, i16 %x11, i32 11
%r12 = insertelement <16 x i16> %r11, i16 %x12, i32 12
%r13 = insertelement <16 x i16> %r12, i16 %x13, i32 13
%r14 = insertelement <16 x i16> %r13, i16 %x14, i32 14
%r15 = insertelement <16 x i16> %r14, i16 %x15, i32 15
ret <16 x i16> %r15
}
define <16 x i16> @var_shuffle_v16i16_v8i16_xxxxxxxxxxxxxxxx_i16(<8 x i16> %x, i32 %i0, i32 %i1, i32 %i2, i32 %i3, i32 %i4, i32 %i5, i32 %i6, i32 %i7, i32 %i8, i32 %i9, i32 %i10, i32 %i11, i32 %i12, i32 %i13, i32 %i14, i32 %i15) nounwind {
; AVX1-LABEL: var_shuffle_v16i16_v8i16_xxxxxxxxxxxxxxxx_i16:
; AVX1: # BB#0:
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: # kill: %R9D<def> %R9D<kill> %R9<def>
; AVX1-NEXT: # kill: %R8D<def> %R8D<kill> %R8<def>
; AVX1-NEXT: # kill: %ECX<def> %ECX<kill> %RCX<def>
; AVX1-NEXT: # kill: %EDX<def> %EDX<kill> %RDX<def>
; AVX1-NEXT: # kill: %ESI<def> %ESI<kill> %RSI<def>
; AVX1-NEXT: # kill: %EDI<def> %EDI<kill> %RDI<def>
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: movzwl -24(%rsp,%rax,2), %eax
; AVX1-NEXT: vmovd %eax, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $1, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $2, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $3, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $4, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $5, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $6, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $7, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $7, %edi
; AVX1-NEXT: movzwl -24(%rsp,%rdi,2), %eax
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovd %eax, %xmm1
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $7, %esi
; AVX1-NEXT: vpinsrw $1, -24(%rsp,%rsi,2), %xmm1, %xmm1
; AVX1-NEXT: andl $7, %edx
; AVX1-NEXT: vpinsrw $2, -24(%rsp,%rdx,2), %xmm1, %xmm1
; AVX1-NEXT: andl $7, %ecx
; AVX1-NEXT: vpinsrw $3, -24(%rsp,%rcx,2), %xmm1, %xmm1
; AVX1-NEXT: andl $7, %r8d
; AVX1-NEXT: vpinsrw $4, -24(%rsp,%r8,2), %xmm1, %xmm1
; AVX1-NEXT: andl $7, %r9d
; AVX1-NEXT: vpinsrw $5, -24(%rsp,%r9,2), %xmm1, %xmm1
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $6, -24(%rsp,%rax,2), %xmm1, %xmm1
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX1-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpinsrw $7, -24(%rsp,%rax,2), %xmm1, %xmm1
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: var_shuffle_v16i16_v8i16_xxxxxxxxxxxxxxxx_i16:
; AVX2: # BB#0:
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: # kill: %R9D<def> %R9D<kill> %R9<def>
; AVX2-NEXT: # kill: %R8D<def> %R8D<kill> %R8<def>
; AVX2-NEXT: # kill: %ECX<def> %ECX<kill> %RCX<def>
; AVX2-NEXT: # kill: %EDX<def> %EDX<kill> %RDX<def>
; AVX2-NEXT: # kill: %ESI<def> %ESI<kill> %RSI<def>
; AVX2-NEXT: # kill: %EDI<def> %EDI<kill> %RDI<def>
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: movzwl -24(%rsp,%rax,2), %eax
; AVX2-NEXT: vmovd %eax, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $1, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $2, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $3, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $4, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $5, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $6, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $7, -24(%rsp,%rax,2), %xmm0, %xmm0
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $7, %edi
; AVX2-NEXT: movzwl -24(%rsp,%rdi,2), %eax
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovd %eax, %xmm1
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $7, %esi
; AVX2-NEXT: vpinsrw $1, -24(%rsp,%rsi,2), %xmm1, %xmm1
; AVX2-NEXT: andl $7, %edx
; AVX2-NEXT: vpinsrw $2, -24(%rsp,%rdx,2), %xmm1, %xmm1
; AVX2-NEXT: andl $7, %ecx
; AVX2-NEXT: vpinsrw $3, -24(%rsp,%rcx,2), %xmm1, %xmm1
; AVX2-NEXT: andl $7, %r8d
; AVX2-NEXT: vpinsrw $4, -24(%rsp,%r8,2), %xmm1, %xmm1
; AVX2-NEXT: andl $7, %r9d
; AVX2-NEXT: vpinsrw $5, -24(%rsp,%r9,2), %xmm1, %xmm1
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $6, -24(%rsp,%rax,2), %xmm1, %xmm1
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: movl {{[0-9]+}}(%rsp), %eax
; AVX2-NEXT: andl $7, %eax
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpinsrw $7, -24(%rsp,%rax,2), %xmm1, %xmm1
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX2-NEXT: retq
%x0 = extractelement <8 x i16> %x, i32 %i0
%x1 = extractelement <8 x i16> %x, i32 %i1
%x2 = extractelement <8 x i16> %x, i32 %i2
%x3 = extractelement <8 x i16> %x, i32 %i3
%x4 = extractelement <8 x i16> %x, i32 %i4
%x5 = extractelement <8 x i16> %x, i32 %i5
%x6 = extractelement <8 x i16> %x, i32 %i6
%x7 = extractelement <8 x i16> %x, i32 %i7
%x8 = extractelement <8 x i16> %x, i32 %i8
%x9 = extractelement <8 x i16> %x, i32 %i9
%x10 = extractelement <8 x i16> %x, i32 %i10
%x11 = extractelement <8 x i16> %x, i32 %i11
%x12 = extractelement <8 x i16> %x, i32 %i12
%x13 = extractelement <8 x i16> %x, i32 %i13
%x14 = extractelement <8 x i16> %x, i32 %i14
%x15 = extractelement <8 x i16> %x, i32 %i15
%r0 = insertelement <16 x i16> undef, i16 %x0 , i32 0
%r1 = insertelement <16 x i16> %r0 , i16 %x1 , i32 1
%r2 = insertelement <16 x i16> %r1 , i16 %x2 , i32 2
%r3 = insertelement <16 x i16> %r2 , i16 %x3 , i32 3
%r4 = insertelement <16 x i16> %r3 , i16 %x4 , i32 4
%r5 = insertelement <16 x i16> %r4 , i16 %x5 , i32 5
%r6 = insertelement <16 x i16> %r5 , i16 %x6 , i32 6
%r7 = insertelement <16 x i16> %r6 , i16 %x7 , i32 7
%r8 = insertelement <16 x i16> %r7 , i16 %x8 , i32 8
%r9 = insertelement <16 x i16> %r8 , i16 %x9 , i32 9
%r10 = insertelement <16 x i16> %r9 , i16 %x10, i32 10
%r11 = insertelement <16 x i16> %r10, i16 %x11, i32 11
%r12 = insertelement <16 x i16> %r11, i16 %x12, i32 12
%r13 = insertelement <16 x i16> %r12, i16 %x13, i32 13
%r14 = insertelement <16 x i16> %r13, i16 %x14, i32 14
%r15 = insertelement <16 x i16> %r14, i16 %x15, i32 15
ret <16 x i16> %r15
}
;
; Unary shuffle indices from memory
;
define <4 x i64> @mem_shuffle_v4i64_v4i64_xxxx_i64(<4 x i64> %x, i64* %i) nounwind {
; AVX1-LABEL: mem_shuffle_v4i64_v4i64_xxxx_i64:
; AVX1: # BB#0:
; AVX1-NEXT: pushq %rbp
; AVX1-NEXT: movq %rsp, %rbp
; AVX1-NEXT: andq $-32, %rsp
; AVX1-NEXT: subq $64, %rsp
; AVX1-NEXT: movq (%rdi), %rax
; AVX1-NEXT: movq 8(%rdi), %rcx
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $3, %eax
; AVX1-NEXT: andl $3, %ecx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: movq 16(%rdi), %rdx
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $3, %edx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: movq 24(%rdi), %rsi
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $3, %esi
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovaps %ymm0, (%rsp)
; AVX1-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX1-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX1-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: movq %rbp, %rsp
; AVX1-NEXT: popq %rbp
; AVX1-NEXT: retq
;
; AVX2-LABEL: mem_shuffle_v4i64_v4i64_xxxx_i64:
; AVX2: # BB#0:
; AVX2-NEXT: pushq %rbp
; AVX2-NEXT: movq %rsp, %rbp
; AVX2-NEXT: andq $-32, %rsp
; AVX2-NEXT: subq $64, %rsp
; AVX2-NEXT: movq (%rdi), %rax
; AVX2-NEXT: movq 8(%rdi), %rcx
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $3, %eax
; AVX2-NEXT: andl $3, %ecx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: movq 16(%rdi), %rdx
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $3, %edx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: movq 24(%rdi), %rsi
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $3, %esi
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovaps %ymm0, (%rsp)
; AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: movq %rbp, %rsp
; AVX2-NEXT: popq %rbp
; AVX2-NEXT: retq
%p0 = getelementptr inbounds i64, i64* %i, i32 0
%p1 = getelementptr inbounds i64, i64* %i, i32 1
%p2 = getelementptr inbounds i64, i64* %i, i32 2
%p3 = getelementptr inbounds i64, i64* %i, i32 3
%i0 = load i64, i64* %p0, align 4
%i1 = load i64, i64* %p1, align 4
%i2 = load i64, i64* %p2, align 4
%i3 = load i64, i64* %p3, align 4
%x0 = extractelement <4 x i64> %x, i64 %i0
%x1 = extractelement <4 x i64> %x, i64 %i1
%x2 = extractelement <4 x i64> %x, i64 %i2
%x3 = extractelement <4 x i64> %x, i64 %i3
%r0 = insertelement <4 x i64> undef, i64 %x0, i32 0
%r1 = insertelement <4 x i64> %r0, i64 %x1, i32 1
%r2 = insertelement <4 x i64> %r1, i64 %x2, i32 2
%r3 = insertelement <4 x i64> %r2, i64 %x3, i32 3
ret <4 x i64> %r3
}
define <4 x i64> @mem_shuffle_v4i64_v2i64_xxxx_i64(<2 x i64> %x, i64* %i) nounwind {
; AVX1-LABEL: mem_shuffle_v4i64_v2i64_xxxx_i64:
; AVX1: # BB#0:
; AVX1-NEXT: movq (%rdi), %rax
; AVX1-NEXT: movq 8(%rdi), %rcx
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $1, %eax
; AVX1-NEXT: andl $1, %ecx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: movq 16(%rdi), %rdx
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $1, %edx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: movq 24(%rdi), %rsi
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX1-NEXT: andl $1, %esi
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; AVX1-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX1-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX1-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX1-NEXT: retq
;
; AVX2-LABEL: mem_shuffle_v4i64_v2i64_xxxx_i64:
; AVX2: # BB#0:
; AVX2-NEXT: movq (%rdi), %rax
; AVX2-NEXT: movq 8(%rdi), %rcx
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $1, %eax
; AVX2-NEXT: andl $1, %ecx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: movq 16(%rdi), %rdx
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $1, %edx
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: movq 24(%rdi), %rsi
DAG: Avoid OOB when legalizing vector indexing If a vector index is out of bounds, the result is supposed to be undefined but is not undefined behavior. Change the legalization for indexing the vector on the stack so that an out of bounds index does not create an out of bounds memory access. llvm-svn: 291604
2017-01-11 06:02:30 +08:00
; AVX2-NEXT: andl $1, %esi
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled. Recommiting with compiler time improvements Recommitting after fixup of 32-bit aliasing sign offset bug in DAGCombiner. * Simplify Consecutive Merge Store Candidate Search Now that address aliasing is much less conservative, push through simplified store merging search and chain alias analysis which only checks for parallel stores through the chain subgraph. This is cleaner as the separation of non-interfering loads/stores from the store-merging logic. When merging stores search up the chain through a single load, and finds all possible stores by looking down from through a load and a TokenFactor to all stores visited. This improves the quality of the output SelectionDAG and the output Codegen (save perhaps for some ARM cases where we correctly constructs wider loads, but then promotes them to float operations which appear but requires more expensive constant generation). Some minor peephole optimizations to deal with improved SubDAG shapes (listed below) Additional Minor Changes: 1. Finishes removing unused AliasLoad code 2. Unifies the chain aggregation in the merged stores across code paths 3. Re-add the Store node to the worklist after calling SimplifyDemandedBits. 4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is arbitrary, but seems sufficient to not cause regressions in tests. 5. Remove Chain dependencies of Memory operations on CopyfromReg nodes as these are captured by data dependence 6. Forward loads-store values through tokenfactors containing {CopyToReg,CopyFromReg} Values. 7. Peephole to convert buildvector of extract_vector_elt to extract_subvector if possible (see CodeGen/AArch64/store-merge.ll) 8. Store merging for the ARM target is restricted to 32-bit as some in some contexts invalid 64-bit operations are being generated. This can be removed once appropriate checks are added. This finishes the change Matt Arsenault started in r246307 and jyknight's original patch. Many tests required some changes as memory operations are now reorderable, improving load-store forwarding. One test in particular is worth noting: CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store forwarding converts a load-store pair into a parallel store and a memory-realized bitcast of the same value. However, because we lose the sharing of the explicit and implicit store values we must create another local store. A similar transformation happens before SelectionDAG as well. Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle llvm-svn: 297695
2017-03-14 08:34:14 +08:00
; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
[X86][SSE] Some basic tests for variable shuffles We don't really support non-constant shuffle masks, but these tests are for cases where BUILD_VECTOR is made up from vector extracts (as well as undef/zero scalars). llvm-svn: 265045
2016-04-01 04:26:30 +08:00
; AVX2-NEXT: retq
%p0 = getelementptr inbounds i64, i64* %i, i32 0
%p1 = getelementptr inbounds i64, i64* %i, i32 1
%p2 = getelementptr inbounds i64, i64* %i, i32 2
%p3 = getelementptr inbounds i64, i64* %i, i32 3
%i0 = load i64, i64* %p0, align 4
%i1 = load i64, i64* %p1, align 4
%i2 = load i64, i64* %p2, align 4
%i3 = load i64, i64* %p3, align 4
%x0 = extractelement <2 x i64> %x, i64 %i0
%x1 = extractelement <2 x i64> %x, i64 %i1
%x2 = extractelement <2 x i64> %x, i64 %i2
%x3 = extractelement <2 x i64> %x, i64 %i3
%r0 = insertelement <4 x i64> undef, i64 %x0, i32 0
%r1 = insertelement <4 x i64> %r0, i64 %x1, i32 1
%r2 = insertelement <4 x i64> %r1, i64 %x2, i32 2
%r3 = insertelement <4 x i64> %r2, i64 %x3, i32 3
ret <4 x i64> %r3
}