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/AArch64/funnel-shift-rot.ll

228 lines
6.5 KiB
LLVM
Raw Normal View History

[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=aarch64-- | FileCheck %s
declare i8 @llvm.fshl.i8(i8, i8, i8)
declare i16 @llvm.fshl.i16(i16, i16, i16)
declare i32 @llvm.fshl.i32(i32, i32, i32)
declare i64 @llvm.fshl.i64(i64, i64, i64)
declare <4 x i32> @llvm.fshl.v4i32(<4 x i32>, <4 x i32>, <4 x i32>)
declare i8 @llvm.fshr.i8(i8, i8, i8)
declare i16 @llvm.fshr.i16(i16, i16, i16)
declare i32 @llvm.fshr.i32(i32, i32, i32)
declare i64 @llvm.fshr.i64(i64, i64, i64)
declare <4 x i32> @llvm.fshr.v4i32(<4 x i32>, <4 x i32>, <4 x i32>)
; When first 2 operands match, it's a rotate.
define i8 @rotl_i8_const_shift(i8 %x) {
; CHECK-LABEL: rotl_i8_const_shift:
; CHECK: // %bb.0:
; CHECK-NEXT: ubfx w8, w0, #5, #3
; CHECK-NEXT: bfi w8, w0, #3, #29
; CHECK-NEXT: mov w0, w8
; CHECK-NEXT: ret
%f = call i8 @llvm.fshl.i8(i8 %x, i8 %x, i8 3)
ret i8 %f
}
define i64 @rotl_i64_const_shift(i64 %x) {
; CHECK-LABEL: rotl_i64_const_shift:
; CHECK: // %bb.0:
; CHECK-NEXT: ror x0, x0, #61
; CHECK-NEXT: ret
%f = call i64 @llvm.fshl.i64(i64 %x, i64 %x, i64 3)
ret i64 %f
}
; When first 2 operands match, it's a rotate (by variable amount).
define i16 @rotl_i16(i16 %x, i16 %z) {
; CHECK-LABEL: rotl_i16:
; CHECK: // %bb.0:
[SelectionDAG] fix bug in translating funnel shift with non-power-of-2 type The bug is visible in the constant-folded x86 tests. We can't use the negated shift amount when the type is not power-of-2: https://rise4fun.com/Alive/US1r ...so in that case, use the regular lowering that includes a select to guard against a shift-by-bitwidth. This path is improved by only calculating the modulo shift amount once now. Also, improve the rotate (with power-of-2 size) lowering to use a negate rather than subtract from bitwidth. This improves the codegen whether we have a rotate instruction or not (although we can still see that we're not matching to a legal rotate in all cases). llvm-svn: 338592
2018-08-02 01:17:08 +08:00
; CHECK-NEXT: neg w10, w1
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; CHECK-NEXT: and w8, w0, #0xffff
; CHECK-NEXT: and w9, w1, #0xf
; CHECK-NEXT: and w10, w10, #0xf
; CHECK-NEXT: lsl w9, w0, w9
; CHECK-NEXT: lsr w8, w8, w10
; CHECK-NEXT: orr w0, w9, w8
; CHECK-NEXT: ret
%f = call i16 @llvm.fshl.i16(i16 %x, i16 %x, i16 %z)
ret i16 %f
}
define i32 @rotl_i32(i32 %x, i32 %z) {
; CHECK-LABEL: rotl_i32:
; CHECK: // %bb.0:
[SelectionDAG] fix bug in translating funnel shift with non-power-of-2 type The bug is visible in the constant-folded x86 tests. We can't use the negated shift amount when the type is not power-of-2: https://rise4fun.com/Alive/US1r ...so in that case, use the regular lowering that includes a select to guard against a shift-by-bitwidth. This path is improved by only calculating the modulo shift amount once now. Also, improve the rotate (with power-of-2 size) lowering to use a negate rather than subtract from bitwidth. This improves the codegen whether we have a rotate instruction or not (although we can still see that we're not matching to a legal rotate in all cases). llvm-svn: 338592
2018-08-02 01:17:08 +08:00
; CHECK-NEXT: neg w8, w1
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; CHECK-NEXT: ror w0, w0, w8
; CHECK-NEXT: ret
%f = call i32 @llvm.fshl.i32(i32 %x, i32 %x, i32 %z)
ret i32 %f
}
[AArch, PowerPC] add more tests for legal rotate ops; NFC llvm-svn: 337964
2018-07-26 05:25:50 +08:00
define i64 @rotl_i64(i64 %x, i64 %z) {
; CHECK-LABEL: rotl_i64:
; CHECK: // %bb.0:
[SelectionDAG] try harder to convert funnel shift to rotate Similar to rL337966 - if the DAGCombiner's rotate matching was working as expected, I don't think we'd see any test diffs here. AArch only goes right, and PPC only goes left. x86 has both, so no diffs there. Differential Revision: https://reviews.llvm.org/D50091 llvm-svn: 339359
2018-08-10 01:26:22 +08:00
; CHECK-NEXT: neg x8, x1
; CHECK-NEXT: ror x0, x0, x8
[AArch, PowerPC] add more tests for legal rotate ops; NFC llvm-svn: 337964
2018-07-26 05:25:50 +08:00
; CHECK-NEXT: ret
%f = call i64 @llvm.fshl.i64(i64 %x, i64 %x, i64 %z)
ret i64 %f
}
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; Vector rotate.
define <4 x i32> @rotl_v4i32(<4 x i32> %x, <4 x i32> %z) {
; CHECK-LABEL: rotl_v4i32:
; CHECK: // %bb.0:
; CHECK-NEXT: movi v2.4s, #31
[SelectionDAG] fix bug in translating funnel shift with non-power-of-2 type The bug is visible in the constant-folded x86 tests. We can't use the negated shift amount when the type is not power-of-2: https://rise4fun.com/Alive/US1r ...so in that case, use the regular lowering that includes a select to guard against a shift-by-bitwidth. This path is improved by only calculating the modulo shift amount once now. Also, improve the rotate (with power-of-2 size) lowering to use a negate rather than subtract from bitwidth. This improves the codegen whether we have a rotate instruction or not (although we can still see that we're not matching to a legal rotate in all cases). llvm-svn: 338592
2018-08-02 01:17:08 +08:00
; CHECK-NEXT: neg v3.4s, v1.4s
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; CHECK-NEXT: and v1.16b, v1.16b, v2.16b
[SelectionDAG] fix bug in translating funnel shift with non-power-of-2 type The bug is visible in the constant-folded x86 tests. We can't use the negated shift amount when the type is not power-of-2: https://rise4fun.com/Alive/US1r ...so in that case, use the regular lowering that includes a select to guard against a shift-by-bitwidth. This path is improved by only calculating the modulo shift amount once now. Also, improve the rotate (with power-of-2 size) lowering to use a negate rather than subtract from bitwidth. This improves the codegen whether we have a rotate instruction or not (although we can still see that we're not matching to a legal rotate in all cases). llvm-svn: 338592
2018-08-02 01:17:08 +08:00
; CHECK-NEXT: and v2.16b, v3.16b, v2.16b
; CHECK-NEXT: neg v2.4s, v2.4s
; CHECK-NEXT: ushl v1.4s, v0.4s, v1.4s
; CHECK-NEXT: ushl v0.4s, v0.4s, v2.4s
; CHECK-NEXT: orr v0.16b, v1.16b, v0.16b
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; CHECK-NEXT: ret
%f = call <4 x i32> @llvm.fshl.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> %z)
ret <4 x i32> %f
}
; Vector rotate by constant splat amount.
define <4 x i32> @rotl_v4i32_rotl_const_shift(<4 x i32> %x) {
; CHECK-LABEL: rotl_v4i32_rotl_const_shift:
; CHECK: // %bb.0:
; CHECK-NEXT: ushr v1.4s, v0.4s, #29
; CHECK-NEXT: shl v0.4s, v0.4s, #3
; CHECK-NEXT: orr v0.16b, v0.16b, v1.16b
; CHECK-NEXT: ret
%f = call <4 x i32> @llvm.fshl.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> <i32 3, i32 3, i32 3, i32 3>)
ret <4 x i32> %f
}
; Repeat everything for funnel shift right.
; When first 2 operands match, it's a rotate.
define i8 @rotr_i8_const_shift(i8 %x) {
; CHECK-LABEL: rotr_i8_const_shift:
; CHECK: // %bb.0:
; CHECK-NEXT: ubfx w8, w0, #3, #5
; CHECK-NEXT: bfi w8, w0, #5, #27
; CHECK-NEXT: mov w0, w8
; CHECK-NEXT: ret
%f = call i8 @llvm.fshr.i8(i8 %x, i8 %x, i8 3)
ret i8 %f
}
define i32 @rotr_i32_const_shift(i32 %x) {
; CHECK-LABEL: rotr_i32_const_shift:
; CHECK: // %bb.0:
; CHECK-NEXT: ror w0, w0, #3
; CHECK-NEXT: ret
%f = call i32 @llvm.fshr.i32(i32 %x, i32 %x, i32 3)
ret i32 %f
}
; When first 2 operands match, it's a rotate (by variable amount).
define i16 @rotr_i16(i16 %x, i16 %z) {
; CHECK-LABEL: rotr_i16:
; CHECK: // %bb.0:
; CHECK-NEXT: and w8, w0, #0xffff
; CHECK-NEXT: and w9, w1, #0xf
[SelectionDAG] fix bug in translating funnel shift with non-power-of-2 type The bug is visible in the constant-folded x86 tests. We can't use the negated shift amount when the type is not power-of-2: https://rise4fun.com/Alive/US1r ...so in that case, use the regular lowering that includes a select to guard against a shift-by-bitwidth. This path is improved by only calculating the modulo shift amount once now. Also, improve the rotate (with power-of-2 size) lowering to use a negate rather than subtract from bitwidth. This improves the codegen whether we have a rotate instruction or not (although we can still see that we're not matching to a legal rotate in all cases). llvm-svn: 338592
2018-08-02 01:17:08 +08:00
; CHECK-NEXT: neg w10, w1
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; CHECK-NEXT: lsr w8, w8, w9
[SelectionDAG] fix bug in translating funnel shift with non-power-of-2 type The bug is visible in the constant-folded x86 tests. We can't use the negated shift amount when the type is not power-of-2: https://rise4fun.com/Alive/US1r ...so in that case, use the regular lowering that includes a select to guard against a shift-by-bitwidth. This path is improved by only calculating the modulo shift amount once now. Also, improve the rotate (with power-of-2 size) lowering to use a negate rather than subtract from bitwidth. This improves the codegen whether we have a rotate instruction or not (although we can still see that we're not matching to a legal rotate in all cases). llvm-svn: 338592
2018-08-02 01:17:08 +08:00
; CHECK-NEXT: and w9, w10, #0xf
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; CHECK-NEXT: lsl w9, w0, w9
; CHECK-NEXT: orr w0, w9, w8
; CHECK-NEXT: ret
%f = call i16 @llvm.fshr.i16(i16 %x, i16 %x, i16 %z)
ret i16 %f
}
[AArch, PowerPC] add more tests for legal rotate ops; NFC llvm-svn: 337964
2018-07-26 05:25:50 +08:00
define i32 @rotr_i32(i32 %x, i32 %z) {
; CHECK-LABEL: rotr_i32:
; CHECK: // %bb.0:
; CHECK-NEXT: ror w0, w0, w1
; CHECK-NEXT: ret
%f = call i32 @llvm.fshr.i32(i32 %x, i32 %x, i32 %z)
ret i32 %f
}
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
define i64 @rotr_i64(i64 %x, i64 %z) {
; CHECK-LABEL: rotr_i64:
; CHECK: // %bb.0:
[SelectionDAG] try to convert funnel shift directly to rotate if legal If the DAGCombiner's rotate matching was working as expected, I don't think we'd see any test diffs here. This sidesteps the issue of custom lowering for rotates raised in PR38243: https://bugs.llvm.org/show_bug.cgi?id=38243 ...by only dealing with legal operations. llvm-svn: 337966
2018-07-26 05:38:30 +08:00
; CHECK-NEXT: ror x0, x0, x1
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; CHECK-NEXT: ret
%f = call i64 @llvm.fshr.i64(i64 %x, i64 %x, i64 %z)
ret i64 %f
}
; Vector rotate.
define <4 x i32> @rotr_v4i32(<4 x i32> %x, <4 x i32> %z) {
; CHECK-LABEL: rotr_v4i32:
; CHECK: // %bb.0:
; CHECK-NEXT: movi v2.4s, #31
[SelectionDAG] fix bug in translating funnel shift with non-power-of-2 type The bug is visible in the constant-folded x86 tests. We can't use the negated shift amount when the type is not power-of-2: https://rise4fun.com/Alive/US1r ...so in that case, use the regular lowering that includes a select to guard against a shift-by-bitwidth. This path is improved by only calculating the modulo shift amount once now. Also, improve the rotate (with power-of-2 size) lowering to use a negate rather than subtract from bitwidth. This improves the codegen whether we have a rotate instruction or not (although we can still see that we're not matching to a legal rotate in all cases). llvm-svn: 338592
2018-08-02 01:17:08 +08:00
; CHECK-NEXT: neg v3.4s, v1.4s
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; CHECK-NEXT: and v1.16b, v1.16b, v2.16b
[SelectionDAG] fix bug in translating funnel shift with non-power-of-2 type The bug is visible in the constant-folded x86 tests. We can't use the negated shift amount when the type is not power-of-2: https://rise4fun.com/Alive/US1r ...so in that case, use the regular lowering that includes a select to guard against a shift-by-bitwidth. This path is improved by only calculating the modulo shift amount once now. Also, improve the rotate (with power-of-2 size) lowering to use a negate rather than subtract from bitwidth. This improves the codegen whether we have a rotate instruction or not (although we can still see that we're not matching to a legal rotate in all cases). llvm-svn: 338592
2018-08-02 01:17:08 +08:00
; CHECK-NEXT: and v2.16b, v3.16b, v2.16b
; CHECK-NEXT: neg v1.4s, v1.4s
; CHECK-NEXT: ushl v1.4s, v0.4s, v1.4s
; CHECK-NEXT: ushl v0.4s, v0.4s, v2.4s
; CHECK-NEXT: orr v0.16b, v0.16b, v1.16b
[Intrinsics] define funnel shift IR intrinsics + DAG builder support As discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.html http://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions, and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift" operation which may also be a single hardware instruction. Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working on that (much larger patch), but then I concluded that we don't need it. At least as a first step, we have all of the backend support necessary to match these ops...because it was required. And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are likely all that we'll ever need. There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes (along with improving the oversized shift documentation). Again, I don't think that's strictly necessary (as the test results here prove). That can be an efficiency improvement as a small follow-up patch. So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support. Differential Revision: https://reviews.llvm.org/D49242 llvm-svn: 337221
2018-07-17 06:59:31 +08:00
; CHECK-NEXT: ret
%f = call <4 x i32> @llvm.fshr.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> %z)
ret <4 x i32> %f
}
; Vector rotate by constant splat amount.
define <4 x i32> @rotr_v4i32_const_shift(<4 x i32> %x) {
; CHECK-LABEL: rotr_v4i32_const_shift:
; CHECK: // %bb.0:
; CHECK-NEXT: ushr v1.4s, v0.4s, #3
; CHECK-NEXT: shl v0.4s, v0.4s, #29
; CHECK-NEXT: orr v0.16b, v0.16b, v1.16b
; CHECK-NEXT: ret
%f = call <4 x i32> @llvm.fshr.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> <i32 3, i32 3, i32 3, i32 3>)
ret <4 x i32> %f
}
define i32 @rotl_i32_shift_by_bitwidth(i32 %x) {
; CHECK-LABEL: rotl_i32_shift_by_bitwidth:
; CHECK: // %bb.0:
; CHECK-NEXT: ret
%f = call i32 @llvm.fshl.i32(i32 %x, i32 %x, i32 32)
ret i32 %f
}
define i32 @rotr_i32_shift_by_bitwidth(i32 %x) {
; CHECK-LABEL: rotr_i32_shift_by_bitwidth:
; CHECK: // %bb.0:
; CHECK-NEXT: ret
%f = call i32 @llvm.fshr.i32(i32 %x, i32 %x, i32 32)
ret i32 %f
}
define <4 x i32> @rotl_v4i32_shift_by_bitwidth(<4 x i32> %x) {
; CHECK-LABEL: rotl_v4i32_shift_by_bitwidth:
; CHECK: // %bb.0:
; CHECK-NEXT: ret
%f = call <4 x i32> @llvm.fshl.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> <i32 32, i32 32, i32 32, i32 32>)
ret <4 x i32> %f
}
define <4 x i32> @rotr_v4i32_shift_by_bitwidth(<4 x i32> %x) {
; CHECK-LABEL: rotr_v4i32_shift_by_bitwidth:
; CHECK: // %bb.0:
; CHECK-NEXT: ret
%f = call <4 x i32> @llvm.fshr.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> <i32 32, i32 32, i32 32, i32 32>)
ret <4 x i32> %f
}