From 57a55313c3dcea2f3208d9d16463908d8eb41603 Mon Sep 17 00:00:00 2001 From: Matt Arsenault Date: Fri, 3 Apr 2020 11:09:19 -0400 Subject: [PATCH] InstCombine: Reduce minnum/maxnum if inputs are casted --- .../InstCombine/InstCombineCalls.cpp | 16 +++++ .../InstCombine/double-float-shrink-1.ll | 12 ++-- llvm/test/Transforms/InstCombine/maximum.ll | 24 ++++++++ llvm/test/Transforms/InstCombine/maxnum.ll | 58 +++++++++++++++++++ llvm/test/Transforms/InstCombine/minimum.ll | 54 +++++++++++++---- llvm/test/Transforms/InstCombine/minnum.ll | 57 ++++++++++++++++++ 6 files changed, 204 insertions(+), 17 deletions(-) diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp index 4dbe517c54298..cd0b5e87e5f11 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp @@ -2316,6 +2316,22 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { } } + Value *ExtSrc0; + Value *ExtSrc1; + + // minnum (fpext x), (fpext y) -> minnum x, y + // maxnum (fpext x), (fpext y) -> maxnum x, y + if (match(II->getArgOperand(0), m_OneUse(m_FPExt(m_Value(ExtSrc0)))) && + match(II->getArgOperand(1), m_OneUse(m_FPExt(m_Value(ExtSrc1)))) && + ExtSrc0->getType() == ExtSrc1->getType()) { + Value *F = Intrinsic::getDeclaration(II->getModule(), II->getIntrinsicID(), + { ExtSrc0->getType() }); + CallInst *NewCall = Builder.CreateCall(F, { ExtSrc0, ExtSrc1 }); + NewCall->copyFastMathFlags(II); + NewCall->takeName(II); + return new FPExtInst(NewCall, II->getType()); + } + break; } case Intrinsic::fmuladd: { diff --git a/llvm/test/Transforms/InstCombine/double-float-shrink-1.ll b/llvm/test/Transforms/InstCombine/double-float-shrink-1.ll index 330e0abb36b85..65439eb1429f8 100644 --- a/llvm/test/Transforms/InstCombine/double-float-shrink-1.ll +++ b/llvm/test/Transforms/InstCombine/double-float-shrink-1.ll @@ -529,12 +529,14 @@ define float @max1(float %a, float %b) { define float @fake_fmin(float %a, float %b) { ; CHECK-LABEL: @fake_fmin( -; CHECK-NEXT: [[C:%.*]] = fpext float [[A:%.*]] to fp128 -; CHECK-NEXT: [[D:%.*]] = fpext float [[B:%.*]] to fp128 -; ISC99-NEXT: [[E:%.*]] = call nsz fp128 @llvm.minnum.f128(fp128 [[C]], fp128 [[D]]) +; ISC99-NEXT: [[MIN:%.*]] = call nsz float @llvm.minnum.f32(float %a, float %b) +; ISC99-NEXT: ret float [[MIN]] + +; ISC89-NEXT: [[C:%.*]] = fpext float [[A:%.*]] to fp128 +; ISC89-NEXT: [[D:%.*]] = fpext float [[B:%.*]] to fp128 ; ISC89-NEXT: [[E:%.*]] = call fp128 @fmin(fp128 [[C]], fp128 [[D]]) -; CHECK-NEXT: [[F:%.*]] = fptrunc fp128 [[E]] to float -; CHECK-NEXT: ret float [[F]] +; ISC89-NEXT: [[F:%.*]] = fptrunc fp128 [[E]] to float +; ISC89-NEXT: ret float [[F]] ; %c = fpext float %a to fp128 %d = fpext float %b to fp128 diff --git a/llvm/test/Transforms/InstCombine/maximum.ll b/llvm/test/Transforms/InstCombine/maximum.ll index a60d0a6685afa..2439f7cd5f502 100644 --- a/llvm/test/Transforms/InstCombine/maximum.ll +++ b/llvm/test/Transforms/InstCombine/maximum.ll @@ -361,3 +361,27 @@ define float @unary_neg_neg_extra_use_x_and_y(float %x, float %y) { call void @use(float %negy) ret float %r } + +define float @reduce_precision(float %x, float %y) { +; CHECK-LABEL: @reduce_precision( +; CHECK-NEXT: [[MAXIMUM:%.*]] = call float @llvm.maximum.f32(float [[X:%.*]], float [[Y:%.*]]) +; CHECK-NEXT: ret float [[MAXIMUM]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + %maximum = call double @llvm.maximum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %maximum to float + ret float %trunc +} + +define float @reduce_precision_fmf(float %x, float %y) { +; CHECK-LABEL: @reduce_precision_fmf( +; CHECK-NEXT: [[MAXIMUM:%.*]] = call nnan float @llvm.maximum.f32(float [[X:%.*]], float [[Y:%.*]]) +; CHECK-NEXT: ret float [[MAXIMUM]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + %maximum = call nnan double @llvm.maximum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %maximum to float + ret float %trunc +} diff --git a/llvm/test/Transforms/InstCombine/maxnum.ll b/llvm/test/Transforms/InstCombine/maxnum.ll index c89e30598db44..df3c75c7e53f6 100644 --- a/llvm/test/Transforms/InstCombine/maxnum.ll +++ b/llvm/test/Transforms/InstCombine/maxnum.ll @@ -355,3 +355,61 @@ define float @unary_neg_neg_extra_use_x_and_y(float %x, float %y) { call void @use(float %negy) ret float %r } + +define float @reduce_precision(float %x, float %y) { +; CHECK-LABEL: @reduce_precision( +; CHECK-NEXT: [[MAXNUM:%.*]] = call float @llvm.maxnum.f32(float [[X:%.*]], float [[Y:%.*]]) +; CHECK-NEXT: ret float [[MAXNUM]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + %maxnum = call double @llvm.maxnum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %maxnum to float + ret float %trunc +} + +define float @reduce_precision_fmf(float %x, float %y) { +; CHECK-LABEL: @reduce_precision_fmf( +; CHECK-NEXT: [[MAXNUM:%.*]] = call nnan float @llvm.maxnum.f32(float [[X:%.*]], float [[Y:%.*]]) +; CHECK-NEXT: ret float [[MAXNUM]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + %maxnum = call nnan double @llvm.maxnum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %maxnum to float + ret float %trunc +} + +define float @reduce_precision_multi_use_0(float %x, float %y) { +; CHECK-LABEL: @reduce_precision_multi_use_0( +; CHECK-NEXT: [[X_EXT:%.*]] = fpext float [[X:%.*]] to double +; CHECK-NEXT: [[Y_EXT:%.*]] = fpext float [[Y:%.*]] to double +; CHECK-NEXT: store double [[X_EXT]], double* undef, align 8 +; CHECK-NEXT: [[MAXNUM:%.*]] = call double @llvm.maxnum.f64(double [[X_EXT]], double [[Y_EXT]]) +; CHECK-NEXT: [[TRUNC:%.*]] = fptrunc double [[MAXNUM]] to float +; CHECK-NEXT: ret float [[TRUNC]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + store double %x.ext, double* undef + %maxnum = call double @llvm.maxnum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %maxnum to float + ret float %trunc +} + +define float @reduce_precision_multi_use_1(float %x, float %y) { +; CHECK-LABEL: @reduce_precision_multi_use_1( +; CHECK-NEXT: [[X_EXT:%.*]] = fpext float [[X:%.*]] to double +; CHECK-NEXT: [[Y_EXT:%.*]] = fpext float [[Y:%.*]] to double +; CHECK-NEXT: store double [[Y_EXT]], double* undef, align 8 +; CHECK-NEXT: [[MAXNUM:%.*]] = call double @llvm.maxnum.f64(double [[X_EXT]], double [[Y_EXT]]) +; CHECK-NEXT: [[TRUNC:%.*]] = fptrunc double [[MAXNUM]] to float +; CHECK-NEXT: ret float [[TRUNC]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + store double %y.ext, double* undef + %maxnum = call double @llvm.maxnum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %maxnum to float + ret float %trunc +} diff --git a/llvm/test/Transforms/InstCombine/minimum.ll b/llvm/test/Transforms/InstCombine/minimum.ll index 158e44ce9cc53..a8b6aefe52c49 100644 --- a/llvm/test/Transforms/InstCombine/minimum.ll +++ b/llvm/test/Transforms/InstCombine/minimum.ll @@ -149,8 +149,9 @@ define float @minimum_f32_val_nan(float %x) { define float @minimum_f32_1_minimum_val_p0(float %x) { ; CHECK-LABEL: @minimum_f32_1_minimum_val_p0( -; CHECK-NEXT: [[RES:%.*]] = call float @llvm.minimum.f32(float %x, float 0.000000e+00) -; CHECK-NEXT: ret float [[RES]] +; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.minimum.f32(float [[X:%.*]], float 0.000000e+00) +; CHECK-NEXT: ret float [[TMP1]] +; %y = call float @llvm.minimum.f32(float %x, float 0.0) %z = call float @llvm.minimum.f32(float %y, float 1.0) ret float %z @@ -158,8 +159,9 @@ define float @minimum_f32_1_minimum_val_p0(float %x) { define float @minimum_f32_1_minimum_p0_val_fast(float %x) { ; CHECK-LABEL: @minimum_f32_1_minimum_p0_val_fast( -; CHECK-NEXT: [[RES:%.*]] = call fast float @llvm.minimum.f32(float %x, float 0.000000e+00) -; CHECK-NEXT: ret float [[RES]] +; CHECK-NEXT: [[TMP1:%.*]] = call fast float @llvm.minimum.f32(float [[X:%.*]], float 0.000000e+00) +; CHECK-NEXT: ret float [[TMP1]] +; %y = call float @llvm.minimum.f32(float 0.0, float %x) %z = call fast float @llvm.minimum.f32(float %y, float 1.0) ret float %z @@ -167,8 +169,9 @@ define float @minimum_f32_1_minimum_p0_val_fast(float %x) { define float @minimum_f32_1_minimum_p0_val_nnan_ninf(float %x) { ; CHECK-LABEL: @minimum_f32_1_minimum_p0_val_nnan_ninf( -; CHECK-NEXT: [[RES:%.*]] = call nnan ninf float @llvm.minimum.f32(float %x, float 0.000000e+00) -; CHECK-NEXT: ret float [[RES]] +; CHECK-NEXT: [[TMP1:%.*]] = call nnan ninf float @llvm.minimum.f32(float [[X:%.*]], float 0.000000e+00) +; CHECK-NEXT: ret float [[TMP1]] +; %y = call float @llvm.minimum.f32(float 0.0, float %x) %z = call nnan ninf float @llvm.minimum.f32(float %y, float 1.0) ret float %z @@ -176,8 +179,9 @@ define float @minimum_f32_1_minimum_p0_val_nnan_ninf(float %x) { define float @minimum_f32_p0_minimum_val_n0(float %x) { ; CHECK-LABEL: @minimum_f32_p0_minimum_val_n0( -; CHECK-NEXT: [[RES:%.*]] = call float @llvm.minimum.f32(float %x, float -0.000000e+00) -; CHECK-NEXT: ret float [[RES]] +; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.minimum.f32(float [[X:%.*]], float -0.000000e+00) +; CHECK-NEXT: ret float [[TMP1]] +; %y = call float @llvm.minimum.f32(float %x, float -0.0) %z = call float @llvm.minimum.f32(float %y, float 0.0) ret float %z @@ -185,8 +189,9 @@ define float @minimum_f32_p0_minimum_val_n0(float %x) { define float @minimum_f32_1_minimum_p0_val(float %x) { ; CHECK-LABEL: @minimum_f32_1_minimum_p0_val( -; CHECK-NEXT: [[RES:%.*]] = call float @llvm.minimum.f32(float %x, float 0.000000e+00) -; CHECK-NEXT: ret float [[RES]] +; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.minimum.f32(float [[X:%.*]], float 0.000000e+00) +; CHECK-NEXT: ret float [[TMP1]] +; %y = call float @llvm.minimum.f32(float 0.0, float %x) %z = call float @llvm.minimum.f32(float %y, float 1.0) ret float %z @@ -194,8 +199,9 @@ define float @minimum_f32_1_minimum_p0_val(float %x) { define <2 x float> @minimum_f32_1_minimum_val_p0_val_v2f32(<2 x float> %x) { ; CHECK-LABEL: @minimum_f32_1_minimum_val_p0_val_v2f32( -; CHECK-NEXT: [[RES:%.*]] = call <2 x float> @llvm.minimum.v2f32(<2 x float> %x, <2 x float> zeroinitializer) -; CHECK-NEXT: ret <2 x float> [[RES]] +; CHECK-NEXT: [[TMP1:%.*]] = call <2 x float> @llvm.minimum.v2f32(<2 x float> [[X:%.*]], <2 x float> zeroinitializer) +; CHECK-NEXT: ret <2 x float> [[TMP1]] +; %y = call <2 x float> @llvm.minimum.v2f32(<2 x float> %x, <2 x float> zeroinitializer) %z = call <2 x float> @llvm.minimum.v2f32(<2 x float> %y, <2 x float>) ret <2 x float> %z @@ -386,3 +392,27 @@ define double @unary_neg_neg_extra_use_x_and_y(double %x, double %y) { call void @use(double %negy) ret double %r } + +define float @reduce_precision(float %x, float %y) { +; CHECK-LABEL: @reduce_precision( +; CHECK-NEXT: [[MINIMUM:%.*]] = call float @llvm.minimum.f32(float [[X:%.*]], float [[Y:%.*]]) +; CHECK-NEXT: ret float [[MINIMUM]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + %minimum = call double @llvm.minimum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %minimum to float + ret float %trunc +} + +define float @reduce_precision_fmf(float %x, float %y) { +; CHECK-LABEL: @reduce_precision_fmf( +; CHECK-NEXT: [[MINIMUM:%.*]] = call nnan float @llvm.minimum.f32(float [[X:%.*]], float [[Y:%.*]]) +; CHECK-NEXT: ret float [[MINIMUM]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + %minimum = call nnan double @llvm.minimum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %minimum to float + ret float %trunc +} diff --git a/llvm/test/Transforms/InstCombine/minnum.ll b/llvm/test/Transforms/InstCombine/minnum.ll index f65d055c66e63..1c3e5a803c0df 100644 --- a/llvm/test/Transforms/InstCombine/minnum.ll +++ b/llvm/test/Transforms/InstCombine/minnum.ll @@ -393,3 +393,60 @@ define double @unary_neg_neg_extra_use_x_and_y(double %x, double %y) { ret double %r } +define float @reduce_precision(float %x, float %y) { +; CHECK-LABEL: @reduce_precision( +; CHECK-NEXT: [[MINNUM:%.*]] = call float @llvm.minnum.f32(float [[X:%.*]], float [[Y:%.*]]) +; CHECK-NEXT: ret float [[MINNUM]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + %minnum = call double @llvm.minnum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %minnum to float + ret float %trunc +} + +define float @reduce_precision_fmf(float %x, float %y) { +; CHECK-LABEL: @reduce_precision_fmf( +; CHECK-NEXT: [[MINNUM:%.*]] = call nnan float @llvm.minnum.f32(float [[X:%.*]], float [[Y:%.*]]) +; CHECK-NEXT: ret float [[MINNUM]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + %minnum = call nnan double @llvm.minnum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %minnum to float + ret float %trunc +} + +define float @reduce_precision_multi_use_0(float %x, float %y) { +; CHECK-LABEL: @reduce_precision_multi_use_0( +; CHECK-NEXT: [[X_EXT:%.*]] = fpext float [[X:%.*]] to double +; CHECK-NEXT: [[Y_EXT:%.*]] = fpext float [[Y:%.*]] to double +; CHECK-NEXT: store double [[X_EXT]], double* undef, align 8 +; CHECK-NEXT: [[MINNUM:%.*]] = call double @llvm.minnum.f64(double [[X_EXT]], double [[Y_EXT]]) +; CHECK-NEXT: [[TRUNC:%.*]] = fptrunc double [[MINNUM]] to float +; CHECK-NEXT: ret float [[TRUNC]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + store double %x.ext, double* undef + %minnum = call double @llvm.minnum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %minnum to float + ret float %trunc +} + +define float @reduce_precision_multi_use_1(float %x, float %y) { +; CHECK-LABEL: @reduce_precision_multi_use_1( +; CHECK-NEXT: [[X_EXT:%.*]] = fpext float [[X:%.*]] to double +; CHECK-NEXT: [[Y_EXT:%.*]] = fpext float [[Y:%.*]] to double +; CHECK-NEXT: store double [[Y_EXT]], double* undef, align 8 +; CHECK-NEXT: [[MINNUM:%.*]] = call double @llvm.minnum.f64(double [[X_EXT]], double [[Y_EXT]]) +; CHECK-NEXT: [[TRUNC:%.*]] = fptrunc double [[MINNUM]] to float +; CHECK-NEXT: ret float [[TRUNC]] +; + %x.ext = fpext float %x to double + %y.ext = fpext float %y to double + store double %y.ext, double* undef + %minnum = call double @llvm.minnum.f64(double %x.ext, double %y.ext) + %trunc = fptrunc double %minnum to float + ret float %trunc +}