[PATCH] D66016: [DAGCombiner] exclude x*2.0 from normal negation profitability rules
Sanjay Patel via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 9 08:40:17 PDT 2019
spatel created this revision.
spatel added reviewers: efriedma, mcberg2017, craig.topper, cameron.mcinally.
Herald added subscribers: hiraditya, kristof.beyls, javed.absar, mcrosier.
Herald added a project: LLVM.
This is the codegen part of fixing:
https://bugs.llvm.org/show_bug.cgi?id=32939
Even with the optimal/canonical IR that is ideally created by D65954 <https://reviews.llvm.org/D65954>, we would reverse that transform in DAGCombiner and end up with the same asm on AArch64 or x86.
I see 2 options for trying to correct this:
1. Limit isNegatibleForFree() by special-casing the fmul pattern (this patch).
2. Avoid creating (fmul X, 2.0) in the 1st place by adding a special-case transform to SelectionDAG::getNode() that matches the transform done by DAGCombiner.
This seems like the less intrusive patch, but let me know if there's some other reason to prefer 1 option over the other.
https://reviews.llvm.org/D66016
Files:
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
llvm/test/CodeGen/AArch64/fadd-combines.ll
llvm/test/CodeGen/X86/fadd-combines.ll
Index: llvm/test/CodeGen/X86/fadd-combines.ll
===================================================================
--- llvm/test/CodeGen/X86/fadd-combines.ll
+++ llvm/test/CodeGen/X86/fadd-combines.ll
@@ -252,9 +252,9 @@
define double @fmul2_negated(double %a, double %b, double %c) {
; CHECK-LABEL: fmul2_negated:
; CHECK: # %bb.0:
-; CHECK-NEXT: mulsd {{.*}}(%rip), %xmm1
+; CHECK-NEXT: addsd %xmm1, %xmm1
; CHECK-NEXT: mulsd %xmm2, %xmm1
-; CHECK-NEXT: addsd %xmm1, %xmm0
+; CHECK-NEXT: subsd %xmm1, %xmm0
; CHECK-NEXT: retq
%mul = fmul double %b, 2.0
%mul1 = fmul double %mul, %c
@@ -265,9 +265,9 @@
define <2 x double> @fmul2_negated_vec(<2 x double> %a, <2 x double> %b, <2 x double> %c) {
; CHECK-LABEL: fmul2_negated_vec:
; CHECK: # %bb.0:
-; CHECK-NEXT: mulpd {{.*}}(%rip), %xmm1
+; CHECK-NEXT: addpd %xmm1, %xmm1
; CHECK-NEXT: mulpd %xmm2, %xmm1
-; CHECK-NEXT: addpd %xmm1, %xmm0
+; CHECK-NEXT: subpd %xmm1, %xmm0
; CHECK-NEXT: retq
%mul = fmul <2 x double> %b, <double 2.0, double 2.0>
%mul1 = fmul <2 x double> %mul, %c
Index: llvm/test/CodeGen/AArch64/fadd-combines.ll
===================================================================
--- llvm/test/CodeGen/AArch64/fadd-combines.ll
+++ llvm/test/CodeGen/AArch64/fadd-combines.ll
@@ -169,10 +169,9 @@
define double @fmul2_negated(double %a, double %b, double %c) {
; CHECK-LABEL: fmul2_negated:
; CHECK: // %bb.0:
-; CHECK-NEXT: fmov d3, #-2.00000000
-; CHECK-NEXT: fmul d1, d1, d3
+; CHECK-NEXT: fadd d1, d1, d1
; CHECK-NEXT: fmul d1, d1, d2
-; CHECK-NEXT: fadd d0, d0, d1
+; CHECK-NEXT: fsub d0, d0, d1
; CHECK-NEXT: ret
%mul = fmul double %b, 2.0
%mul1 = fmul double %mul, %c
@@ -183,10 +182,9 @@
define <2 x double> @fmul2_negated_vec(<2 x double> %a, <2 x double> %b, <2 x double> %c) {
; CHECK-LABEL: fmul2_negated_vec:
; CHECK: // %bb.0:
-; CHECK-NEXT: fmov v3.2d, #-2.00000000
-; CHECK-NEXT: fmul v1.2d, v1.2d, v3.2d
+; CHECK-NEXT: fadd v1.2d, v1.2d, v1.2d
; CHECK-NEXT: fmul v1.2d, v1.2d, v2.2d
-; CHECK-NEXT: fadd v0.2d, v0.2d, v1.2d
+; CHECK-NEXT: fsub v0.2d, v0.2d, v1.2d
; CHECK-NEXT: ret
%mul = fmul <2 x double> %b, <double 2.0, double 2.0>
%mul1 = fmul <2 x double> %mul, %c
Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -868,6 +868,11 @@
Options, ForCodeSize, Depth + 1))
return V;
+ // Ignore X * 2.0 because that is expected to be canonicalized to X + X.
+ if (auto *C = isConstOrConstSplatFP(Op.getOperand(1)))
+ if (C->isExactlyValue(2.0) && Op.getOpcode() == ISD::FMUL)
+ return 0;
+
return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
ForCodeSize, Depth + 1);
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D66016.214382.patch
Type: text/x-patch
Size: 2995 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20190809/af141de0/attachment.bin>
More information about the llvm-commits
mailing list