2016-09-20 22:42:45 +08:00
|
|
|
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
|
|
; RUN: llc < %s -mtriple=i386-unknown-linux-gnu -mattr=sse2 | FileCheck %s
|
2015-10-23 00:14:45 +08:00
|
|
|
|
|
|
|
; Source file looks something like this:
|
|
|
|
;
|
|
|
|
; typedef int AAA[100][100];
|
|
|
|
;
|
|
|
|
; void testCombineMultiplies(AAA a,int lll)
|
|
|
|
; {
|
|
|
|
; int LOC = lll + 5;
|
|
|
|
;
|
|
|
|
; a[LOC][LOC] = 11;
|
|
|
|
;
|
|
|
|
; a[LOC][20] = 22;
|
|
|
|
; a[LOC+20][20] = 33;
|
|
|
|
; }
|
|
|
|
;
|
|
|
|
; We want to make sure we don't generate 2 multiply instructions,
|
|
|
|
; one for a[LOC][] and one for a[LOC+20]. visitMUL in DAGCombiner.cpp
|
|
|
|
; should combine the instructions in such a way to avoid the extra
|
|
|
|
; multiply.
|
|
|
|
;
|
|
|
|
; Output looks roughly like this:
|
|
|
|
;
|
|
|
|
; movl 8(%esp), %eax
|
|
|
|
; movl 12(%esp), %ecx
|
|
|
|
; imull $400, %ecx, %edx # imm = 0x190
|
|
|
|
; leal (%edx,%eax), %esi
|
|
|
|
; movl $11, 2020(%esi,%ecx,4)
|
|
|
|
; movl $22, 2080(%edx,%eax)
|
|
|
|
; movl $33, 10080(%edx,%eax)
|
|
|
|
|
|
|
|
; Function Attrs: nounwind
|
2016-09-20 22:42:45 +08:00
|
|
|
define void @testCombineMultiplies([100 x i32]* nocapture %a, i32 %lll) nounwind {
|
|
|
|
; CHECK-LABEL: testCombineMultiplies:
|
2017-12-05 01:18:51 +08:00
|
|
|
; CHECK: # %bb.0: # %entry
|
2016-09-20 22:42:45 +08:00
|
|
|
; CHECK-NEXT: pushl %esi
|
|
|
|
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
|
|
|
|
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx
|
|
|
|
; CHECK-NEXT: imull $400, %ecx, %edx # imm = 0x190
|
[DAG] Refactor DAGCombiner::ReassociateOps
Summary:
Extract the logic for doing reassociations
from DAGCombiner::reassociateOps into a helper
function DAGCombiner::reassociateOpsCommutative,
and use that helper to trigger reassociation
on the original operand order, or the commuted
operand order.
Codegen is not identical since the operand order will
be different when doing the reassociations for the
commuted case. That causes some unfortunate churn in
some test cases. Apart from that this should be NFC.
Reviewers: spatel, craig.topper, tstellar
Reviewed By: spatel
Subscribers: dmgreen, dschuff, jvesely, nhaehnle, javed.absar, sbc100, jgravelle-google, hiraditya, aheejin, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61199
llvm-svn: 359476
2019-04-30 01:50:10 +08:00
|
|
|
; CHECK-NEXT: leal (%edx,%eax), %esi
|
2016-09-20 22:42:45 +08:00
|
|
|
; CHECK-NEXT: movl $11, 2020(%esi,%ecx,4)
|
[DAG] Refactor DAGCombiner::ReassociateOps
Summary:
Extract the logic for doing reassociations
from DAGCombiner::reassociateOps into a helper
function DAGCombiner::reassociateOpsCommutative,
and use that helper to trigger reassociation
on the original operand order, or the commuted
operand order.
Codegen is not identical since the operand order will
be different when doing the reassociations for the
commuted case. That causes some unfortunate churn in
some test cases. Apart from that this should be NFC.
Reviewers: spatel, craig.topper, tstellar
Reviewed By: spatel
Subscribers: dmgreen, dschuff, jvesely, nhaehnle, javed.absar, sbc100, jgravelle-google, hiraditya, aheejin, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61199
llvm-svn: 359476
2019-04-30 01:50:10 +08:00
|
|
|
; CHECK-NEXT: movl $22, 2080(%edx,%eax)
|
|
|
|
; CHECK-NEXT: movl $33, 10080(%edx,%eax)
|
2016-09-20 22:42:45 +08:00
|
|
|
; CHECK-NEXT: popl %esi
|
|
|
|
; CHECK-NEXT: retl
|
2015-10-23 00:14:45 +08:00
|
|
|
entry:
|
|
|
|
%add = add nsw i32 %lll, 5
|
|
|
|
%arrayidx1 = getelementptr inbounds [100 x i32], [100 x i32]* %a, i32 %add, i32 %add
|
|
|
|
store i32 11, i32* %arrayidx1, align 4
|
|
|
|
%arrayidx3 = getelementptr inbounds [100 x i32], [100 x i32]* %a, i32 %add, i32 20
|
|
|
|
store i32 22, i32* %arrayidx3, align 4
|
|
|
|
%add4 = add nsw i32 %lll, 25
|
|
|
|
%arrayidx6 = getelementptr inbounds [100 x i32], [100 x i32]* %a, i32 %add4, i32 20
|
|
|
|
store i32 33, i32* %arrayidx6, align 4
|
|
|
|
ret void
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
; Test for the same optimization on vector multiplies.
|
|
|
|
;
|
|
|
|
; Source looks something like this:
|
|
|
|
;
|
|
|
|
; typedef int v4int __attribute__((__vector_size__(16)));
|
|
|
|
;
|
|
|
|
; v4int x;
|
|
|
|
; v4int v2, v3;
|
|
|
|
; void testCombineMultiplies_splat(v4int v1) {
|
|
|
|
; v2 = (v1 + (v4int){ 11, 11, 11, 11 }) * (v4int) {22, 22, 22, 22};
|
|
|
|
; v3 = (v1 + (v4int){ 33, 33, 33, 33 }) * (v4int) {22, 22, 22, 22};
|
|
|
|
; x = (v1 + (v4int){ 11, 11, 11, 11 });
|
|
|
|
; }
|
|
|
|
;
|
|
|
|
; Output looks something like this:
|
|
|
|
;
|
|
|
|
; testCombineMultiplies_splat: # @testCombineMultiplies_splat
|
2017-12-05 01:18:51 +08:00
|
|
|
; # %bb.0: # %entry
|
2015-10-23 00:14:45 +08:00
|
|
|
; movdqa .LCPI1_0, %xmm1 # xmm1 = [11,11,11,11]
|
|
|
|
; paddd %xmm0, %xmm1
|
|
|
|
; movdqa .LCPI1_1, %xmm2 # xmm2 = [22,22,22,22]
|
|
|
|
; pshufd $245, %xmm0, %xmm3 # xmm3 = xmm0[1,1,3,3]
|
|
|
|
; pmuludq %xmm2, %xmm0
|
|
|
|
; pshufd $232, %xmm0, %xmm0 # xmm0 = xmm0[0,2,2,3]
|
|
|
|
; pmuludq %xmm2, %xmm3
|
|
|
|
; pshufd $232, %xmm3, %xmm2 # xmm2 = xmm3[0,2,2,3]
|
|
|
|
; punpckldq %xmm2, %xmm0 # xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
|
|
|
|
; movdqa .LCPI1_2, %xmm2 # xmm2 = [242,242,242,242]
|
|
|
|
; paddd %xmm0, %xmm2
|
|
|
|
; paddd .LCPI1_3, %xmm0
|
|
|
|
; movdqa %xmm2, v2
|
|
|
|
; movdqa %xmm0, v3
|
|
|
|
; movdqa %xmm1, x
|
|
|
|
; retl
|
|
|
|
;
|
|
|
|
; Again, we want to make sure we don't generate two different multiplies.
|
|
|
|
; We should have a single multiply for "v1 * {22, 22, 22, 22}" (made up of two
|
|
|
|
; pmuludq instructions), followed by two adds. Without this optimization, we'd
|
|
|
|
; do 2 adds, followed by 2 multiplies (i.e. 4 pmuludq instructions).
|
|
|
|
|
|
|
|
@v2 = common global <4 x i32> zeroinitializer, align 16
|
|
|
|
@v3 = common global <4 x i32> zeroinitializer, align 16
|
|
|
|
@x = common global <4 x i32> zeroinitializer, align 16
|
|
|
|
|
|
|
|
; Function Attrs: nounwind
|
2016-09-20 22:42:45 +08:00
|
|
|
define void @testCombineMultiplies_splat(<4 x i32> %v1) nounwind {
|
|
|
|
; CHECK-LABEL: testCombineMultiplies_splat:
|
2017-12-05 01:18:51 +08:00
|
|
|
; CHECK: # %bb.0: # %entry
|
2016-09-20 22:42:45 +08:00
|
|
|
; CHECK-NEXT: movdqa {{.*#+}} xmm1 = [11,11,11,11]
|
|
|
|
; CHECK-NEXT: paddd %xmm0, %xmm1
|
|
|
|
; CHECK-NEXT: movdqa {{.*#+}} xmm2 = [22,22,22,22]
|
|
|
|
; CHECK-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
|
|
|
|
; CHECK-NEXT: pmuludq %xmm2, %xmm0
|
|
|
|
; CHECK-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
|
|
|
|
; CHECK-NEXT: pmuludq %xmm2, %xmm3
|
|
|
|
; CHECK-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,2,2,3]
|
|
|
|
; CHECK-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
|
|
|
|
; CHECK-NEXT: movdqa {{.*#+}} xmm2 = [242,242,242,242]
|
|
|
|
; CHECK-NEXT: paddd %xmm0, %xmm2
|
|
|
|
; CHECK-NEXT: paddd {{\.LCPI.*}}, %xmm0
|
|
|
|
; CHECK-NEXT: movdqa %xmm2, v2
|
|
|
|
; CHECK-NEXT: movdqa %xmm0, v3
|
|
|
|
; CHECK-NEXT: movdqa %xmm1, x
|
|
|
|
; CHECK-NEXT: retl
|
2015-10-23 00:14:45 +08:00
|
|
|
entry:
|
|
|
|
%add1 = add <4 x i32> %v1, <i32 11, i32 11, i32 11, i32 11>
|
|
|
|
%mul1 = mul <4 x i32> %add1, <i32 22, i32 22, i32 22, i32 22>
|
|
|
|
%add2 = add <4 x i32> %v1, <i32 33, i32 33, i32 33, i32 33>
|
|
|
|
%mul2 = mul <4 x i32> %add2, <i32 22, i32 22, i32 22, i32 22>
|
|
|
|
store <4 x i32> %mul1, <4 x i32>* @v2, align 16
|
|
|
|
store <4 x i32> %mul2, <4 x i32>* @v3, align 16
|
|
|
|
store <4 x i32> %add1, <4 x i32>* @x, align 16
|
|
|
|
ret void
|
|
|
|
}
|
|
|
|
|
|
|
|
; Finally, check the non-splatted vector case. This is very similar
|
|
|
|
; to the previous test case, except for the vector values.
|
2016-09-20 22:42:45 +08:00
|
|
|
|
2015-10-23 00:14:45 +08:00
|
|
|
; Function Attrs: nounwind
|
2016-09-20 22:42:45 +08:00
|
|
|
define void @testCombineMultiplies_non_splat(<4 x i32> %v1) nounwind {
|
|
|
|
; CHECK-LABEL: testCombineMultiplies_non_splat:
|
2017-12-05 01:18:51 +08:00
|
|
|
; CHECK: # %bb.0: # %entry
|
2016-09-20 22:42:45 +08:00
|
|
|
; CHECK-NEXT: movdqa {{.*#+}} xmm1 = [11,22,33,44]
|
|
|
|
; CHECK-NEXT: paddd %xmm0, %xmm1
|
|
|
|
; CHECK-NEXT: movdqa {{.*#+}} xmm2 = [22,33,44,55]
|
|
|
|
; CHECK-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
|
|
|
|
; CHECK-NEXT: pmuludq %xmm2, %xmm0
|
|
|
|
; CHECK-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
|
|
|
|
; CHECK-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
|
|
|
|
; CHECK-NEXT: pmuludq %xmm3, %xmm2
|
|
|
|
; CHECK-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
|
|
|
|
; CHECK-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
|
|
|
|
; CHECK-NEXT: movdqa {{.*#+}} xmm2 = [242,726,1452,2420]
|
|
|
|
; CHECK-NEXT: paddd %xmm0, %xmm2
|
|
|
|
; CHECK-NEXT: paddd {{\.LCPI.*}}, %xmm0
|
|
|
|
; CHECK-NEXT: movdqa %xmm2, v2
|
|
|
|
; CHECK-NEXT: movdqa %xmm0, v3
|
|
|
|
; CHECK-NEXT: movdqa %xmm1, x
|
|
|
|
; CHECK-NEXT: retl
|
2015-10-23 00:14:45 +08:00
|
|
|
entry:
|
|
|
|
%add1 = add <4 x i32> %v1, <i32 11, i32 22, i32 33, i32 44>
|
|
|
|
%mul1 = mul <4 x i32> %add1, <i32 22, i32 33, i32 44, i32 55>
|
|
|
|
%add2 = add <4 x i32> %v1, <i32 33, i32 44, i32 55, i32 66>
|
|
|
|
%mul2 = mul <4 x i32> %add2, <i32 22, i32 33, i32 44, i32 55>
|
|
|
|
store <4 x i32> %mul1, <4 x i32>* @v2, align 16
|
|
|
|
store <4 x i32> %mul2, <4 x i32>* @v3, align 16
|
|
|
|
store <4 x i32> %add1, <4 x i32>* @x, align 16
|
|
|
|
ret void
|
|
|
|
}
|