From b40fde062c306584d88243de11175187e754ce3b Mon Sep 17 00:00:00 2001
From: Sanjay Patel <spatel@rotateright.com>
Date: Sun, 14 Feb 2021 08:06:15 -0500
Subject: [PATCH] [InstCombine] fold fdiv with pow divisor (PR49147)

This is unusual in the general (non-reciprocal) case because we need
an extra instruction, but that should be better for general FP
reassociation and codegen. We conservatively check for "arcp" FMF
here as we do with existing fdiv folds, but it is not strictly
necessary to have that.

This is part of solving:
https://llvm.org/PR49147
(The powi variant potentially has a different constraint.)

Differential Revision: https://reviews.llvm.org/D96648
---
 .../Transforms/InstCombine/InstCombineMulDivRem.cpp   | 11 +++++++++++
 llvm/test/Transforms/InstCombine/fdiv.ll              | 19 ++++++++++++++-----
 2 files changed, 25 insertions(+), 5 deletions(-)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
index 5526739..9bc566e 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -1325,6 +1325,17 @@ Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) {
     // replaced by a multiplication.
     if (match(Op1, m_FDiv(m_SpecificFP(1.0), m_Value(Y))))
       return BinaryOperator::CreateFMulFMF(Y, Op0, &I);
+
+    // Negate the exponent of pow to fold division-by-pow() into multiply:
+    // Z / pow(X, Y) --> Z * pow(X, -Y)
+    // In the general case, this creates an extra instruction, but fmul allows
+    // for better canonicalization and optimization than fdiv.
+    if (match(Op1,
+              m_OneUse(m_Intrinsic<Intrinsic::pow>(m_Value(X), m_Value(Y))))) {
+      Value *NegY = Builder.CreateFNegFMF(Y, &I);
+      Value *Pow = Builder.CreateBinaryIntrinsic(Intrinsic::pow, X, NegY, &I);
+      return BinaryOperator::CreateFMulFMF(Op0, Pow, &I);
+    }
   }
 
   if (I.hasAllowReassoc() && Op0->hasOneUse() && Op1->hasOneUse()) {
diff --git a/llvm/test/Transforms/InstCombine/fdiv.ll b/llvm/test/Transforms/InstCombine/fdiv.ll
index f1858cb..52bf746 100644
--- a/llvm/test/Transforms/InstCombine/fdiv.ll
+++ b/llvm/test/Transforms/InstCombine/fdiv.ll
@@ -661,8 +661,9 @@ define float @fabs_fabs_extra_use3(float %x, float %y) {
 
 define float @pow_divisor(float %x, float %y, float %z) {
 ; CHECK-LABEL: @pow_divisor(
-; CHECK-NEXT:    [[P:%.*]] = call float @llvm.pow.f32(float [[X:%.*]], float [[Y:%.*]])
-; CHECK-NEXT:    [[R:%.*]] = fdiv reassoc arcp float [[Z:%.*]], [[P]]
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg reassoc arcp float [[Y:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call reassoc arcp float @llvm.pow.f32(float [[X:%.*]], float [[TMP1]])
+; CHECK-NEXT:    [[R:%.*]] = fmul reassoc arcp float [[TMP2]], [[Z:%.*]]
 ; CHECK-NEXT:    ret float [[R]]
 ;
   %p = call float @llvm.pow.f32(float %x, float %y)
@@ -670,6 +671,8 @@ define float @pow_divisor(float %x, float %y, float %z) {
   ret float %r
 }
 
+; Negative test - don't create an extra pow
+
 define float @pow_divisor_extra_use(float %x, float %y, float %z) {
 ; CHECK-LABEL: @pow_divisor_extra_use(
 ; CHECK-NEXT:    [[P:%.*]] = call float @llvm.pow.f32(float [[X:%.*]], float [[Y:%.*]])
@@ -683,6 +686,8 @@ define float @pow_divisor_extra_use(float %x, float %y, float %z) {
   ret float %r
 }
 
+; Negative test - must have reassoc+arcp
+
 define float @pow_divisor_not_enough_fmf(float %x, float %y, float %z) {
 ; CHECK-LABEL: @pow_divisor_not_enough_fmf(
 ; CHECK-NEXT:    [[P:%.*]] = call fast float @llvm.pow.f32(float [[X:%.*]], float [[Y:%.*]])
@@ -694,6 +699,8 @@ define float @pow_divisor_not_enough_fmf(float %x, float %y, float %z) {
   ret float %r
 }
 
+; Negative test - must have reassoc+arcp
+
 define float @pow_divisor_not_enough_fmf2(float %x, float %y, float %z) {
 ; CHECK-LABEL: @pow_divisor_not_enough_fmf2(
 ; CHECK-NEXT:    [[P:%.*]] = call fast float @llvm.pow.f32(float [[X:%.*]], float [[Y:%.*]])
@@ -705,11 +712,13 @@ define float @pow_divisor_not_enough_fmf2(float %x, float %y, float %z) {
   ret float %r
 }
 
+; Special-case - reciprocal does not require extra fmul
+
 define <2 x half> @pow_recip(<2 x half> %x, <2 x half> %y) {
 ; CHECK-LABEL: @pow_recip(
-; CHECK-NEXT:    [[P:%.*]] = call <2 x half> @llvm.pow.v2f16(<2 x half> [[X:%.*]], <2 x half> [[Y:%.*]])
-; CHECK-NEXT:    [[R:%.*]] = fdiv reassoc ninf arcp <2 x half> <half 0xH3C00, half 0xH3C00>, [[P]]
-; CHECK-NEXT:    ret <2 x half> [[R]]
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg reassoc ninf arcp <2 x half> [[Y:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call reassoc ninf arcp <2 x half> @llvm.pow.v2f16(<2 x half> [[X:%.*]], <2 x half> [[TMP1]])
+; CHECK-NEXT:    ret <2 x half> [[TMP2]]
 ;
   %p = call <2 x half> @llvm.pow.v2f16(<2 x half> %x, <2 x half> %y)
   %r = fdiv reassoc arcp ninf <2 x half> <half 1.0, half 1.0>, %p
-- 
2.7.4