[PATCH] D86403: [DAGCombine]: Fold X/Sqrt(X) to Sqrt(X).

[Venkataramanan Kumar via Phabricator via llvm-commits]

venkataramanan.kumar.llvm created this revision.
venkataramanan.kumar.llvm added a reviewer: spatel.
Herald added subscribers: llvm-commits, ecnelises, hiraditya.
Herald added a project: LLVM.
venkataramanan.kumar.llvm requested review of this revision.

Under -ffast-math  ( "nnan" and " reassoc")  fold X/Sqrt(X) to Sqrt(X).


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D86403

Files:
  llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  llvm/test/CodeGen/X86/sqrt-fastmath.ll


Index: llvm/test/CodeGen/X86/sqrt-fastmath.ll
===================================================================
--- llvm/test/CodeGen/X86/sqrt-fastmath.ll
+++ llvm/test/CodeGen/X86/sqrt-fastmath.ll
@@ -903,14 +903,12 @@
 define double @sqrt_fdiv_common_operand(double %x) nounwind {
 ; SSE-LABEL: sqrt_fdiv_common_operand:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    sqrtsd %xmm0, %xmm1
-; SSE-NEXT:    divsd %xmm1, %xmm0
+; SSE-NEXT:    sqrtsd %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: sqrt_fdiv_common_operand:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vsqrtsd %xmm0, %xmm0, %xmm1
-; AVX-NEXT:    vdivsd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vsqrtsd %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %sqrt = call fast double @llvm.sqrt.f64(double %x)
   %r = fdiv fast double %x, %sqrt
@@ -920,33 +918,29 @@
 define <2 x double> @sqrt_fdiv_common_operand_vec(<2 x double> %x) nounwind {
 ; SSE-LABEL: sqrt_fdiv_common_operand_vec:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    sqrtpd %xmm0, %xmm1
-; SSE-NEXT:    divpd %xmm1, %xmm0
+; SSE-NEXT:    sqrtpd %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: sqrt_fdiv_common_operand_vec:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vsqrtpd %xmm0, %xmm1
-; AVX-NEXT:    vdivpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vsqrtpd %xmm0, %xmm0,%xmm0
 ; AVX-NEXT:    retq
   %sqrt = call <2 x double> @llvm.sqrt.v2f64(<2 x double> %x)
-  %r = fdiv arcp reassoc <2 x double> %x, %sqrt
+  %r = fdiv nnan reassoc <2 x double> %x, %sqrt
   ret <2 x double> %r
 }
 
 define double @sqrt_fdiv_common_operand_extra_use(double %x, double* %p) nounwind {
 ; SSE-LABEL: sqrt_fdiv_common_operand_extra_use:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    sqrtsd %xmm0, %xmm1
-; SSE-NEXT:    movsd %xmm1, (%rdi)
-; SSE-NEXT:    divsd %xmm1, %xmm0
+; SSE-NEXT:    sqrtsd %xmm0, %xmm0
+; SSE-NEXT:    movsd %xmm0, (%rdi)
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: sqrt_fdiv_common_operand_extra_use:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vsqrtsd %xmm0, %xmm0, %xmm1
-; AVX-NEXT:    vmovsd %xmm1, (%rdi)
-; AVX-NEXT:    vdivsd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vsqrtsd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovsd %xmm0, (%rdi)
 ; AVX-NEXT:    retq
   %sqrt = call fast double @llvm.sqrt.f64(double %x)
   store double %sqrt, double* %p
Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -13353,6 +13353,13 @@
         return RV;
   }
 
+  // Fold X/Sqrt(X) -> Sqrt(X)
+  if ((Options.NoNaNsFPMath || Flags.hasNoNaNs()) &&
+      (Options.UnsafeFPMath || Flags.hasAllowReassociation()))
+    if (N1.getOpcode() == ISD::FSQRT && N0 == N1.getOperand(0)) {
+      return N1;
+    }
+
   // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
   TargetLowering::NegatibleCost CostN0 =
       TargetLowering::NegatibleCost::Expensive;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D86403.287207.patch
Type: text/x-patch
Size: 2885 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200822/5bddab0e/attachment.bin>