[PATCH] D86726: [InstCombine]: Transform 1.0/sqrt(X) * X to X/sqrt(X) and X * 1.0/sqrt(X) to X/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, steven.zhang, hiraditya.
Herald added a project: LLVM.
venkataramanan.kumar.llvm requested review of this revision.

The patch does the following transforms in InstCombine.

1.0/sqrt(X) * X => X/sqrt(X) 
 X * 1.0/sqrt(X) => X/sqrt(X)

DagCombiner <https://reviews.llvm.org/rG62e91bf563337fff93be0059ecb1f4dae32d8980> can then optimizes the X/sqrt(x) to sqrt(X).


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D86726

Files:
  llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
  llvm/test/Transforms/InstCombine/fmul-sqrt.ll


Index: llvm/test/Transforms/InstCombine/fmul-sqrt.ll
===================================================================
--- llvm/test/Transforms/InstCombine/fmul-sqrt.ll
+++ llvm/test/Transforms/InstCombine/fmul-sqrt.ll
@@ -99,6 +99,45 @@
   ret double %squared
 }
 
+define double @rsqrt_a_reassociate(double %x) {
+; CHECK-LABEL: @rsqrt_a_reassociate(
+; CHECK-NEXT:    [[SQRT:%.*]] = call fast double @llvm.sqrt.f64(double [[X:%.*]])
+; CHECK-NEXT:    [[MULT:%.*]] = fdiv reassoc double [[X:%.*]], [[SQRT]]
+; CHECK-NEXT:    ret double [[MULT]]
+;
+  %sqrt = call fast double @llvm.sqrt.f64(double %x)
+  %rsqrt = fdiv fast double 1.0, %sqrt
+  %mult = fmul reassoc double %rsqrt, %x
+  ret double %mult
+}
+
+define double @a_rsqrt_reassociate(double %x) {
+; CHECK-LABEL: @a_rsqrt_reassociate(
+; CHECK-NEXT:    [[SQRT:%.*]] = call fast double @llvm.sqrt.f64(double [[X:%.*]])
+; CHECK-NEXT:    [[DIV:%.*]] = fdiv reassoc double [[X:%.*]], [[SQRT]]
+; CHECK-NEXT:    ret double [[DIV]]
+;
+  %sqrt = call fast double @llvm.sqrt.f64(double %x)
+  %rsqrt = fdiv fast double 1.0, %sqrt
+  %mult = fmul reassoc double %x, %rsqrt
+  ret double %mult
+}
+
+define double @rsqrt_a_reassociate_extra_use(double %x, double * %p) {
+; CHECK-LABEL: @rsqrt_a_reassociate_extra_use(
+; CHECK-NEXT:    [[SQRT:%.*]] = call fast double @llvm.sqrt.f64(double [[X:%.*]])
+; CHECK-NEXT:    [[RSQRT:%.*]] = fdiv fast double 1.000000e+00, [[SQRT]]
+; CHECK-NEXT:    [[DIV:%.*]] = fdiv reassoc double [[X:%.*]], [[SQRT]]
+; CHECK-NEXT:    store double [[RSQRT]], double* %p
+; CHECK-NEXT:    ret double [[DIV]]
+;
+  %sqrt = call fast double @llvm.sqrt.f64(double %x)
+  %rsqrt = fdiv fast double 1.0, %sqrt
+  %res = fmul reassoc double %rsqrt, %x
+  store double %rsqrt, double* %p
+  ret double %res
+}
+
 define double @sqrt_divisor_squared(double %x, double %y) {
 ; CHECK-LABEL: @sqrt_divisor_squared(
 ; CHECK-NEXT:    [[TMP1:%.*]] = fmul reassoc nnan nsz double [[Y:%.*]], [[Y]]
Index: llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
===================================================================
--- llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -544,6 +544,16 @@
       return replaceInstUsesWith(I, Sqrt);
     }
 
+    // 1.0/sqrt(X) * X = X/sqrt(X).
+    if (match(Op0, (m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) &&
+        match(Y, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) && Op1 == X)
+      return BinaryOperator::CreateFDivFMF(X, Y, &I);
+
+    // X * 1.0/sqrt(X) = X/sqrt(X).
+    if (match(Op1, (m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) &&
+        match(Y, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) && Op0 == X)
+      return BinaryOperator::CreateFDivFMF(X, Y, &I);
+
     // Like the similar transform in instsimplify, this requires 'nsz' because
     // sqrt(-0.0) = -0.0, and -0.0 * -0.0 does not simplify to -0.0.
     if (I.hasNoNaNs() && I.hasNoSignedZeros() && Op0 == Op1 &&


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D86726.288384.patch
Type: text/x-patch
Size: 2975 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200827/928fccbc/attachment.bin>