Convert fcmp with 0.0 from casted integers to icmp

This is already handled in general when it is known the
conversion can't lose bits with smaller integer types
casted into wider floating point types.

This pattern happens somewhat often in GPU programs that cast
workitem intrinsics to float, which are often compared with 0.

Specifically handle the special case of compares with zero which
should also be known to not lose information. I had a more general
version of this which allows equality compares if the casted float is
exactly representable in the integer, but I'm not 100% confident that
is always correct.

Also fold cases that aren't integers to true / false.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@225265 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Matt Arsenault 2015-01-06 15:50:59 +00:00
parent 23dc3f4eef
commit d883ca0ca7
2 changed files with 488 additions and 4 deletions

View File

@ -12,6 +12,7 @@
//===----------------------------------------------------------------------===//
#include "InstCombine.h"
#include "llvm/ADT/APSInt.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/InstructionSimplify.h"
@ -3624,18 +3625,49 @@ Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
if (MantissaWidth == -1) return nullptr; // Unknown.
IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
// Check to see that the input is converted from an integer type that is small
// enough that preserves all bits. TODO: check here for "known" sign bits.
// This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
unsigned InputSize = IntTy->getScalarSizeInBits();
// If this is a uitofp instruction, we need an extra bit to hold the sign.
bool LHSUnsigned = isa<UIToFPInst>(LHSI);
if (LHSUnsigned)
++InputSize;
if (I.isEquality()) {
FCmpInst::Predicate P = I.getPredicate();
bool IsExact = false;
APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
// If the floating point constant isn't an integer value, we know if we will
// ever compare equal / not equal to it.
if (!IsExact) {
// TODO: Can never be -0.0 and other non-representable values
APFloat RHSRoundInt(RHS);
RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
return ReplaceInstUsesWith(I, Builder->getFalse());
assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
return ReplaceInstUsesWith(I, Builder->getTrue());
}
}
// TODO: If the constant is exactly representable, is it always OK to do
// equality compares as integer?
}
// Comparisons with zero are a special case where we know we won't lose
// information.
bool IsCmpZero = RHS.isPosZero();
// If the conversion would lose info, don't hack on this.
if ((int)InputSize > MantissaWidth)
if ((int)InputSize > MantissaWidth && !IsCmpZero)
return nullptr;
// Otherwise, we can potentially simplify the comparison. We know that it
@ -3676,8 +3708,6 @@ Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
return ReplaceInstUsesWith(I, Builder->getFalse());
}
IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
// Now we know that the APFloat is a normal number, zero or inf.
// See if the FP constant is too large for the integer. For example,

View File

@ -0,0 +1,454 @@
; RUN: opt -S -instcombine < %s | FileCheck %s
; CHECK-LABEL: @i32_cast_cmp_oeq_int_0_uitofp(
; CHECK-NEXT: icmp eq i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp oeq float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_n0_uitofp(
; CHECK: uitofp
; CHECK: fcmp oeq
define i1 @i32_cast_cmp_oeq_int_n0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp oeq float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_0_sitofp(
; CHECK-NEXT: icmp eq i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp oeq float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_n0_sitofp(
; CHECK: sitofp
; CHECK: fcmp oeq
define i1 @i32_cast_cmp_oeq_int_n0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp oeq float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_one_int_0_uitofp(
; CHECK-NEXT: icmp ne i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_one_int_0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp one float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_one_int_n0_uitofp(
; CHECK: uitofp
; CHECK: fcmp one
define i1 @i32_cast_cmp_one_int_n0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp one float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_one_int_0_sitofp(
; CHECK-NEXT: icmp ne i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_one_int_0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp one float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_one_int_n0_sitofp(
; CHECK: sitofp
; CHECK: fcmp one
define i1 @i32_cast_cmp_one_int_n0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp one float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ueq_int_0_uitofp(
; CHECK-NEXT: icmp eq i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_ueq_int_0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp ueq float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ueq_int_n0_uitofp(
; CHECK: uitofp
; CHECK: fcmp ueq
define i1 @i32_cast_cmp_ueq_int_n0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp ueq float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ueq_int_0_sitofp(
; CHECK-NEXT: icmp eq i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_ueq_int_0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp ueq float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ueq_int_n0_sitofp(
; CHECK: sitofp
; CHECK: fcmp ueq
define i1 @i32_cast_cmp_ueq_int_n0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp ueq float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_une_int_0_uitofp(
; CHECK-NEXT: icmp ne i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_une_int_0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp une float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_une_int_n0_uitofp(
; CHECK: uitofp
; CHECK: fcmp une
define i1 @i32_cast_cmp_une_int_n0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp une float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_une_int_0_sitofp(
; CHECK-NEXT: icmp ne i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_une_int_0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp une float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_une_int_n0_sitofp(
; CHECK: sitofp
; CHECK: fcmp une
define i1 @i32_cast_cmp_une_int_n0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp une float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ogt_int_0_uitofp(
; CHECK: icmp ne i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_ogt_int_0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp ogt float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ogt_int_n0_uitofp(
; CHECK: uitofp
; CHECK: fcmp ogt
define i1 @i32_cast_cmp_ogt_int_n0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp ogt float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ogt_int_0_sitofp(
; CHECK: icmp sgt i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_ogt_int_0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp ogt float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ogt_int_n0_sitofp(
; CHECK: sitofp
; CHECK: fcmp ogt
define i1 @i32_cast_cmp_ogt_int_n0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp ogt float %f, -0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ole_int_0_uitofp(
; CHECK: icmp eq i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_ole_int_0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp ole float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ole_int_0_sitofp(
; CHECK: icmp slt i32 %i, 1
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_ole_int_0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp ole float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_olt_int_0_uitofp(
; CHECK: ret i1 false
define i1 @i32_cast_cmp_olt_int_0_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp olt float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_olt_int_0_sitofp(
; CHECK: icmp slt i32 %i, 0
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_olt_int_0_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp olt float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i64_cast_cmp_oeq_int_0_uitofp(
; CHECK-NEXT: icmp eq i64 %i, 0
; CHECK-NEXT: ret
define i1 @i64_cast_cmp_oeq_int_0_uitofp(i64 %i) {
%f = uitofp i64 %i to float
%cmp = fcmp oeq float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i64_cast_cmp_oeq_int_0_sitofp(
; CHECK-NEXT: icmp eq i64 %i, 0
; CHECK-NEXT: ret
define i1 @i64_cast_cmp_oeq_int_0_sitofp(i64 %i) {
%f = sitofp i64 %i to float
%cmp = fcmp oeq float %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i64_cast_cmp_oeq_int_0_uitofp_half(
; CHECK-NEXT: icmp eq i64 %i, 0
; CHECK-NEXT: ret
define i1 @i64_cast_cmp_oeq_int_0_uitofp_half(i64 %i) {
%f = uitofp i64 %i to half
%cmp = fcmp oeq half %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i64_cast_cmp_oeq_int_0_sitofp_half(
; CHECK-NEXT: icmp eq i64 %i, 0
; CHECK-NEXT: ret
define i1 @i64_cast_cmp_oeq_int_0_sitofp_half(i64 %i) {
%f = sitofp i64 %i to half
%cmp = fcmp oeq half %f, 0.0
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_0_uitofp_ppcf128(
; CHECK: uitofp
; CHECK: fcmp oeq
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_0_uitofp_ppcf128(i32 %i) {
%f = uitofp i32 %i to ppc_fp128
%cmp = fcmp oeq ppc_fp128 %f, 0xM00000000000000000000000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i24max_uitofp(
; CHECK: uitofp
; CHECK: fcmp oeq
; XCHECK: icmp eq i32 %i, 16777215
; XCHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i24max_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp oeq float %f, 0x416FFFFFE0000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i24max_sitofp(
; CHECK: sitofp
; CHECK: fcmp oeq
; XCHECK: icmp eq i32 %i, 16777215
; XCHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i24max_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp oeq float %f, 0x416FFFFFE0000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i24maxp1_uitofp(
; CHECK: uitofp
; CHECK: fcmp oeq
; XCHECK: icmp eq i32 %i, 16777216
; XCHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i24maxp1_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp oeq float %f, 0x4170000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i24maxp1_sitofp(
; CHECK: sitofp
; CHECK: fcmp oeq
; XCHECK: icmp eq i32 %i, 16777216
; XCHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i24maxp1_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp oeq float %f, 0x4170000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i32umax_uitofp(
; CHECK: uitofp
; CHECK: fcmp oeq
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i32umax_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp oeq float %f, 0x41F0000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i32umax_sitofp(
; CHECK: sitofp
; CHECK: fcmp oeq
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i32umax_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp oeq float %f, 0x41F0000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i32imin_uitofp(
; CHECK: uitofp
; CHECK: fcmp oeq
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i32imin_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp oeq float %f, 0xC1E0000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i32imin_sitofp(
; CHECK: sitofp
; CHECK: fcmp oeq
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i32imin_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp oeq float %f, 0xC1E0000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i32imax_uitofp(
; CHECK: uitofp
; CHECK: fcmp oeq
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i32imax_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp oeq float %f, 0x41E0000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_i32imax_sitofp(
; CHECK: sitofp
; CHECK: fcmp oeq
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_i32imax_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp oeq float %f, 0x41E0000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_negi32umax_uitofp(
; CHECK: uitofp
; CHECK: fcmp oeq
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_negi32umax_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp oeq float %f, 0xC1F0000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_int_negi32umax_sitofp(
; CHECK: sitofp
; CHECK: fcmp oeq
; CHECK-NEXT: ret
define i1 @i32_cast_cmp_oeq_int_negi32umax_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp oeq float %f, 0xC1F0000000000000
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_half_uitofp(
; CHECK: ret i1 false
define i1 @i32_cast_cmp_oeq_half_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp oeq float %f, 0.5
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_oeq_half_sitofp(
; CHECK: ret i1 false
define i1 @i32_cast_cmp_oeq_half_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp oeq float %f, 0.5
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_one_half_uitofp(
; CHECK: ret i1 true
define i1 @i32_cast_cmp_one_half_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp one float %f, 0.5
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_one_half_sitofp(
; CHECK: ret i1 true
define i1 @i32_cast_cmp_one_half_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp one float %f, 0.5
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ueq_half_uitofp(
; CHECK: ret i1 false
define i1 @i32_cast_cmp_ueq_half_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp ueq float %f, 0.5
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_ueq_half_sitofp(
; CHECK: ret i1 false
define i1 @i32_cast_cmp_ueq_half_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp ueq float %f, 0.5
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_une_half_uitofp(
; CHECK: ret i1 true
define i1 @i32_cast_cmp_une_half_uitofp(i32 %i) {
%f = uitofp i32 %i to float
%cmp = fcmp une float %f, 0.5
ret i1 %cmp
}
; CHECK-LABEL: @i32_cast_cmp_une_half_sitofp(
; CHECK: ret i1 true
define i1 @i32_cast_cmp_une_half_sitofp(i32 %i) {
%f = sitofp i32 %i to float
%cmp = fcmp une float %f, 0.5
ret i1 %cmp
}