From 6b00d409003853e8bdaa8c25c4e1e436b1ed2f28 Mon Sep 17 00:00:00 2001 From: Matt Arsenault Date: Mon, 20 Mar 2017 21:59:24 +0000 Subject: [PATCH] InstCombine: Check source value precision when reducing cast intrinsic Missed this check when porting from the libcall version. llvm-svn: 298312 --- .../InstCombine/InstCombineCasts.cpp | 16 +- .../InstCombine/double-float-shrink-2.ll | 441 ++++++++++++++++-- 2 files changed, 419 insertions(+), 38 deletions(-) diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp index 98335ba45429..e08c301ccdd0 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp @@ -1442,10 +1442,22 @@ Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { case Intrinsic::round: case Intrinsic::nearbyint: case Intrinsic::trunc: { + Value *Src = II->getArgOperand(0); + if (!Src->hasOneUse()) + break; + + // Except for fabs, this transformation requires the input of the unary FP + // operation to be itself an fpext from the type to which we're + // truncating. + if (II->getIntrinsicID() != Intrinsic::fabs) { + FPExtInst *FPExtSrc = dyn_cast(Src); + if (!FPExtSrc || FPExtSrc->getOperand(0)->getType() != CI.getType()) + break; + } + // Do unary FP operation on smaller type. // (fptrunc (fabs x)) -> (fabs (fptrunc x)) - Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0), - CI.getType()); + Value *InnerTrunc = Builder->CreateFPTrunc(Src, CI.getType()); Type *IntrinsicType[] = { CI.getType() }; Function *Overload = Intrinsic::getDeclaration( CI.getModule(), II->getIntrinsicID(), IntrinsicType); diff --git a/llvm/test/Transforms/InstCombine/double-float-shrink-2.ll b/llvm/test/Transforms/InstCombine/double-float-shrink-2.ll index 435bd0b0fc9b..4813614f26cb 100644 --- a/llvm/test/Transforms/InstCombine/double-float-shrink-2.ll +++ b/llvm/test/Transforms/InstCombine/double-float-shrink-2.ll @@ -1,32 +1,9 @@ -; RUN: opt < %s -instcombine -S -mtriple "i386-pc-linux" | FileCheck -check-prefix=DO-SIMPLIFY %s -; RUN: opt < %s -instcombine -S -mtriple "i386-pc-win32" | FileCheck -check-prefix=DONT-SIMPLIFY %s -; RUN: opt < %s -instcombine -S -mtriple "x86_64-pc-win32" | FileCheck -check-prefix=C89-SIMPLIFY %s -; RUN: opt < %s -instcombine -S -mtriple "i386-pc-mingw32" | FileCheck -check-prefix=DO-SIMPLIFY %s -; RUN: opt < %s -instcombine -S -mtriple "x86_64-pc-mingw32" | FileCheck -check-prefix=DO-SIMPLIFY %s -; RUN: opt < %s -instcombine -S -mtriple "sparc-sun-solaris" | FileCheck -check-prefix=DO-SIMPLIFY %s - -; DO-SIMPLIFY: call float @llvm.floor.f32( -; DO-SIMPLIFY: call float @llvm.ceil.f32( -; DO-SIMPLIFY: call float @llvm.round.f32( -; DO-SIMPLIFY: call float @llvm.nearbyint.f32( -; DO-SIMPLIFY: call float @llvm.trunc.f32( -; DO-SIMPLIFY: call float @llvm.fabs.f32( -; DO-SIMPLIFY: call fast float @llvm.fabs.f32( - -; C89-SIMPLIFY: call float @llvm.floor.f32( -; C89-SIMPLIFY: call float @llvm.ceil.f32( -; C89-SIMPLIFY: call double @round( -; C89-SIMPLIFY: call double @nearbyint( - -; DONT-SIMPLIFY: call float @llvm.floor.f32( -; DONT-SIMPLIFY: call float @llvm.ceil.f32( -; DONT-SIMPLIFY: call double @round( -; DONT-SIMPLIFY: call double @nearbyint( -; DONT-SIMPLIFY: call double @trunc( - -; This is replaced with the intrinsic, which does the right thing on -; all platforms. -; DONT-SIMPLIFY: call float @llvm.fabs.f32( +; RUN: opt < %s -instcombine -S -mtriple "i386-pc-linux" | FileCheck -check-prefix=DO-SIMPLIFY -check-prefix=ALL %s +; RUN: opt < %s -instcombine -S -mtriple "i386-pc-win32" | FileCheck -check-prefix=DONT-SIMPLIFY -check-prefix=ALL %s +; RUN: opt < %s -instcombine -S -mtriple "x86_64-pc-win32" | FileCheck -check-prefix=C89-SIMPLIFY -check-prefix=ALL %s +; RUN: opt < %s -instcombine -S -mtriple "i386-pc-mingw32" | FileCheck -check-prefix=DO-SIMPLIFY -check-prefix=ALL %s +; RUN: opt < %s -instcombine -S -mtriple "x86_64-pc-mingw32" | FileCheck -check-prefix=DO-SIMPLIFY -check-prefix=ALL %s +; RUN: opt < %s -instcombine -S -mtriple "sparc-sun-solaris" | FileCheck -check-prefix=DO-SIMPLIFY -check-prefix=ALL %s declare double @floor(double) declare double @ceil(double) @@ -34,9 +11,19 @@ declare double @round(double) declare double @nearbyint(double) declare double @trunc(double) declare double @fabs(double) + +declare double @llvm.floor.f64(double) +declare double @llvm.ceil.f64(double) +declare double @llvm.round.f64(double) +declare double @llvm.nearbyint.f64(double) +declare double @llvm.trunc.f64(double) declare double @llvm.fabs.f64(double) -define float @test_floor(float %C) { +; ALL-LABEL: @test_shrink_libcall_floor( +; DO-SIMPLIFY: call float @llvm.floor.f32( +; C89-SIMPLIFY: call float @llvm.floor.f32( +; DONT-SIMPLIFY: call float @llvm.floor.f32( +define float @test_shrink_libcall_floor(float %C) { %D = fpext float %C to double ; --> floorf %E = call double @floor(double %D) @@ -44,7 +31,11 @@ define float @test_floor(float %C) { ret float %F } -define float @test_ceil(float %C) { +; ALL-LABEL: @test_shrink_libcall_ceil( +; DO-SIMPLIFY: call float @llvm.ceil.f32( +; C89-SIMPLIFY: call float @llvm.ceil.f32( +; DONT-SIMPLIFY: call float @llvm.ceil.f32( +define float @test_shrink_libcall_ceil(float %C) { %D = fpext float %C to double ; --> ceilf %E = call double @ceil(double %D) @@ -52,7 +43,11 @@ define float @test_ceil(float %C) { ret float %F } -define float @test_round(float %C) { +; ALL-LABEL: @test_shrink_libcall_round( +; DO-SIMPLIFY: call float @llvm.round.f32( +; C89-SIMPLIFY: call double @round( +; DONT-SIMPLIFY: call double @round( +define float @test_shrink_libcall_round(float %C) { %D = fpext float %C to double ; --> roundf %E = call double @round(double %D) @@ -60,7 +55,11 @@ define float @test_round(float %C) { ret float %F } -define float @test_nearbyint(float %C) { +; ALL-LABEL: @test_shrink_libcall_nearbyint( +; DO-SIMPLIFY: call float @llvm.nearbyint.f32( +; C89-SIMPLIFY: call double @nearbyint( +; DONT-SIMPLIFY: call double @nearbyint( +define float @test_shrink_libcall_nearbyint(float %C) { %D = fpext float %C to double ; --> nearbyintf %E = call double @nearbyint(double %D) @@ -68,7 +67,10 @@ define float @test_nearbyint(float %C) { ret float %F } -define float @test_trunc(float %C) { +; ALL-LABEL: @test_shrink_libcall_trunc( +; DO-SIMPLIFY: call float @llvm.trunc.f32( +; DONT-SIMPLIFY: call double @trunc( +define float @test_shrink_libcall_trunc(float %C) { %D = fpext float %C to double ; --> truncf %E = call double @trunc(double %D) @@ -76,7 +78,13 @@ define float @test_trunc(float %C) { ret float %F } -define float @test_fabs(float %C) { +; ALL-LABEL: @test_shrink_libcall_fabs( +; DO-SIMPLIFY: call float @llvm.fabs.f32( + +; This is replaced with the intrinsic, which does the right thing on +; all platforms. +; DONT-SIMPLIFY: call float @llvm.fabs.f32( +define float @test_shrink_libcall_fabs(float %C) { %D = fpext float %C to double ; --> fabsf %E = call double @fabs(double %D) @@ -85,10 +93,371 @@ define float @test_fabs(float %C) { } ; Make sure fast math flags are preserved -define float @test_fabs_fast(float %C) { +; ALL-LABEL: @test_shrink_libcall_fabs_fast( +; DO-SIMPLIFY: call fast float @llvm.fabs.f32( +define float @test_shrink_libcall_fabs_fast(float %C) { %D = fpext float %C to double ; --> fabsf %E = call fast double @fabs(double %D) %F = fptrunc double %E to float ret float %F } + +; ALL-LABEL: @test_shrink_intrin_floor( +; ALL: call float @llvm.floor.f32( +define float @test_shrink_intrin_floor(float %C) { + %D = fpext float %C to double + ; --> floorf + %E = call double @llvm.floor.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_ceil( +; ALL: call float @llvm.ceil.f32( +define float @test_shrink_intrin_ceil(float %C) { + %D = fpext float %C to double + ; --> ceilf + %E = call double @llvm.ceil.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_round( +; ALL: call float @llvm.round.f32( +define float @test_shrink_intrin_round(float %C) { + %D = fpext float %C to double + ; --> roundf + %E = call double @llvm.round.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_nearbyint( +; ALL: call float @llvm.nearbyint.f32( +define float @test_shrink_intrin_nearbyint(float %C) { + %D = fpext float %C to double + ; --> nearbyintf + %E = call double @llvm.nearbyint.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_trunc( +; ALL-SIMPLIFY: call float @llvm.trunc.f32( +define float @test_shrink_intrin_trunc(float %C) { + %D = fpext float %C to double + %E = call double @llvm.trunc.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_fabs( +; ALL: call float @llvm.fabs.f32( +define float @test_shrink_intrin_fabs(float %C) { + %D = fpext float %C to double + %E = call double @llvm.fabs.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; Make sure fast math flags are preserved +; ALL-LABEL: @test_shrink_intrin_fabs_fast( +; ALL: call fast float @llvm.fabs.f32( +define float @test_shrink_intrin_fabs_fast(float %C) { + %D = fpext float %C to double + %E = call fast double @llvm.fabs.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_no_shrink_intrin_floor( +; ALL: call double @llvm.floor.f64( +define float @test_no_shrink_intrin_floor(double %D) { + %E = call double @llvm.floor.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_no_shrink_intrin_ceil( +; ALL: call double @llvm.ceil.f64( +define float @test_no_shrink_intrin_ceil(double %D) { + %E = call double @llvm.ceil.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_no_shrink_intrin_round( +; ALL: call double @llvm.round.f64( +define float @test_no_shrink_intrin_round(double %D) { + %E = call double @llvm.round.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_no_shrink_intrin_nearbyint( +; ALL: call double @llvm.nearbyint.f64( +define float @test_no_shrink_intrin_nearbyint(double %D) { + %E = call double @llvm.nearbyint.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_no_shrink_intrin_trunc( +; ALL-SIMPLIFY: call double @llvm.trunc.f64( +define float @test_no_shrink_intrin_trunc(double %D) { + %E = call double @llvm.trunc.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_fabs_double_src( +; ALL: call float @llvm.fabs.f32( +define float @test_shrink_intrin_fabs_double_src(double %D) { + %E = call double @llvm.fabs.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; Make sure fast math flags are preserved +; ALL-LABEL: @test_shrink_intrin_fabs_fast_double_src( +; ALL: call fast float @llvm.fabs.f32( +define float @test_shrink_intrin_fabs_fast_double_src(double %D) { + %E = call fast double @llvm.fabs.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_float_convertible_constant_intrin_floor( +; ALL: ret float 2.000000e+00 +define float @test_shrink_float_convertible_constant_intrin_floor() { + %E = call double @llvm.floor.f64(double 2.1) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_float_convertible_constant_intrin_ceil( +; ALL: ret float 3.000000e+00 +define float @test_shrink_float_convertible_constant_intrin_ceil() { + %E = call double @llvm.ceil.f64(double 2.1) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_float_convertible_constant_intrin_round( +; ALL: ret float 2.000000e+00 +define float @test_shrink_float_convertible_constant_intrin_round() { + %E = call double @llvm.round.f64(double 2.1) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_float_convertible_constant_intrin_nearbyint( +; ALL: ret float 2.000000e+00 +define float @test_shrink_float_convertible_constant_intrin_nearbyint() { + %E = call double @llvm.nearbyint.f64(double 2.1) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_float_convertible_constant_intrin_trunc( +; ALL: ret float 2.000000e+00 +define float @test_shrink_float_convertible_constant_intrin_trunc() { + %E = call double @llvm.trunc.f64(double 2.1) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_float_convertible_constant_intrin_fabs( +; ALL: ret float 0x4000CCCCC0000000 +define float @test_shrink_float_convertible_constant_intrin_fabs() { + %E = call double @llvm.fabs.f64(double 2.1) + %F = fptrunc double %E to float + ret float %F +} + +; Make sure fast math flags are preserved +; ALL-LABEL: @test_shrink_float_convertible_constant_intrin_fabs_fast( +; ALL: ret float 0x4000CCCCC0000000 +define float @test_shrink_float_convertible_constant_intrin_fabs_fast() { + %E = call fast double @llvm.fabs.f64(double 2.1) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_no_shrink_mismatched_type_intrin_floor( +; ALL-NEXT: %E = call double @llvm.floor.f64(double %D) +; ALL-NEXT: %F = fptrunc double %E to half +; ALL-NEXT: ret half %F +define half @test_no_shrink_mismatched_type_intrin_floor(double %D) { + %E = call double @llvm.floor.f64(double %D) + %F = fptrunc double %E to half + ret half %F +} + +; ALL-LABEL: @test_no_shrink_mismatched_type_intrin_ceil( +; ALL-NEXT: %E = call double @llvm.ceil.f64(double %D) +; ALL-NEXT: %F = fptrunc double %E to half +; ALL-NEXT: ret half %F +define half @test_no_shrink_mismatched_type_intrin_ceil(double %D) { + %E = call double @llvm.ceil.f64(double %D) + %F = fptrunc double %E to half + ret half %F +} + +; ALL-LABEL: @test_no_shrink_mismatched_type_intrin_round( +; ALL-NEXT: %E = call double @llvm.round.f64(double %D) +; ALL-NEXT: %F = fptrunc double %E to half +; ALL-NEXT: ret half %F +define half @test_no_shrink_mismatched_type_intrin_round(double %D) { + %E = call double @llvm.round.f64(double %D) + %F = fptrunc double %E to half + ret half %F +} + +; ALL-LABEL: @test_no_shrink_mismatched_type_intrin_nearbyint( +; ALL-NEXT: %E = call double @llvm.nearbyint.f64(double %D) +; ALL-NEXT: %F = fptrunc double %E to half +; ALL-NEXT: ret half %F +define half @test_no_shrink_mismatched_type_intrin_nearbyint(double %D) { + %E = call double @llvm.nearbyint.f64(double %D) + %F = fptrunc double %E to half + ret half %F +} + +; ALL-LABEL: @test_no_shrink_mismatched_type_intrin_trunc( +; ALL-NEXT: %E = call double @llvm.trunc.f64(double %D) +; ALL-NEXT: %F = fptrunc double %E to half +; ALL-NEXT: ret half %F +define half @test_no_shrink_mismatched_type_intrin_trunc(double %D) { + %E = call double @llvm.trunc.f64(double %D) + %F = fptrunc double %E to half + ret half %F +} + +; ALL-LABEL: @test_shrink_mismatched_type_intrin_fabs_double_src( +; ALL-NEXT: %1 = fptrunc double %D to half +; ALL-NEXT: %F = call half @llvm.fabs.f16(half %1) +; ALL-NEXT: ret half %F +define half @test_shrink_mismatched_type_intrin_fabs_double_src(double %D) { + %E = call double @llvm.fabs.f64(double %D) + %F = fptrunc double %E to half + ret half %F +} + +; Make sure fast math flags are preserved +; ALL-LABEL: @test_mismatched_type_intrin_fabs_fast_double_src( +; ALL-NEXT: %1 = fptrunc double %D to half +; ALL-NEXT: %F = call fast half @llvm.fabs.f16(half %1) +; ALL-NEXT: ret half %F +define half @test_mismatched_type_intrin_fabs_fast_double_src(double %D) { + %E = call fast double @llvm.fabs.f64(double %D) + %F = fptrunc double %E to half + ret half %F +} + +; ALL-LABEL: @test_shrink_intrin_floor_fp16_src( +; ALL-NEXT: %E = call half @llvm.floor.f16(half %C) +; ALL-NEXT: %1 = fpext half %E to double +; ALL-NEXT: %F = fptrunc double %1 to float +define float @test_shrink_intrin_floor_fp16_src(half %C) { + %D = fpext half %C to double + %E = call double @llvm.floor.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_ceil_fp16_src( +; ALL-NEXT: %E = call half @llvm.ceil.f16(half %C) +; ALL-NEXT: %1 = fpext half %E to double +; ALL-NEXT: %F = fptrunc double %1 to float +; ALL-NEXT: ret float %F +define float @test_shrink_intrin_ceil_fp16_src(half %C) { + %D = fpext half %C to double + %E = call double @llvm.ceil.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_round_fp16_src( +; ALL-NEXT: %E = call half @llvm.round.f16(half %C) +; ALL-NEXT: %1 = fpext half %E to double +; ALL-NEXT: %F = fptrunc double %1 to float +; ALL-NEXT: ret float %F +define float @test_shrink_intrin_round_fp16_src(half %C) { + %D = fpext half %C to double + %E = call double @llvm.round.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_nearbyint_fp16_src( +; ALL-NEXT: %E = call half @llvm.nearbyint.f16(half %C) +; ALL-NEXT: %1 = fpext half %E to double +; ALL-NEXT: %F = fptrunc double %1 to float +; ALL-NEXT: ret float %F +define float @test_shrink_intrin_nearbyint_fp16_src(half %C) { + %D = fpext half %C to double + %E = call double @llvm.nearbyint.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_trunc_fp16_src( +; ALL-NEXT: %E = call half @llvm.trunc.f16(half %C) +; ALL-NEXT: %1 = fpext half %E to double +; ALL-NEXT: %F = fptrunc double %1 to float +; ALL-NEXT: ret float %F +define float @test_shrink_intrin_trunc_fp16_src(half %C) { + %D = fpext half %C to double + %E = call double @llvm.trunc.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_shrink_intrin_fabs_fp16_src( +; ALL-NEXT: %E = call half @llvm.fabs.f16(half %C) +; ALL-NEXT: %1 = fpext half %E to double +; ALL-NEXT: %F = fptrunc double %1 to float +; ALL-NEXT: ret float %F +define float @test_shrink_intrin_fabs_fp16_src(half %C) { + %D = fpext half %C to double + %E = call double @llvm.fabs.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; Make sure fast math flags are preserved +; ALL-LABEL: @test_shrink_intrin_fabs_fast_fp16_src( +; ALL-NEXT: %E = call fast half @llvm.fabs.f16(half %C) +; ALL-NEXT: %1 = fpext half %E to double +; ALL-NEXT: %F = fptrunc double %1 to float +; ALL-NEXT: ret float %F +define float @test_shrink_intrin_fabs_fast_fp16_src(half %C) { + %D = fpext half %C to double + %E = call fast double @llvm.fabs.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_no_shrink_intrin_floor_multi_use_fpext( +; ALL: %D = fpext half %C to double +; ALL: call double @llvm.floor.f64 +define float @test_no_shrink_intrin_floor_multi_use_fpext(half %C) { + %D = fpext half %C to double + store volatile double %D, double* undef + %E = call double @llvm.floor.f64(double %D) + %F = fptrunc double %E to float + ret float %F +} + +; ALL-LABEL: @test_no_shrink_intrin_fabs_multi_use_fpext( +; ALL: %D = fpext half %C to double +; ALL: call double @llvm.fabs.f64 +define float @test_no_shrink_intrin_fabs_multi_use_fpext(half %C) { + %D = fpext half %C to double + store volatile double %D, double* undef + %E = call double @llvm.fabs.f64(double %D) + %F = fptrunc double %E to float + ret float %F +}