[llvm-branch-commits] [llvm] DAG: Handle lowering unordered compare with inf (PR #100378)
Matt Arsenault via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Aug 28 21:19:20 PDT 2024
https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/100378
>From 4edffb2750e8320c39109cd7c9c086c2ee86e9d4 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Tue, 7 Feb 2023 12:22:05 -0400
Subject: [PATCH 1/3] DAG: Handle lowering unordered compare with inf
Try to take advantage of the nan check behavior of fcmp.
x86_64 looks better, x86_32 looks worse.
---
llvm/include/llvm/CodeGen/CodeGenCommonISel.h | 7 +-
llvm/lib/CodeGen/CodeGenCommonISel.cpp | 8 +-
.../CodeGen/SelectionDAG/TargetLowering.cpp | 53 +++++++------
llvm/test/CodeGen/X86/is_fpclass.ll | 78 +++++++++----------
4 files changed, 83 insertions(+), 63 deletions(-)
diff --git a/llvm/include/llvm/CodeGen/CodeGenCommonISel.h b/llvm/include/llvm/CodeGen/CodeGenCommonISel.h
index 90ef890f22d1b1..e4b2e20babc07a 100644
--- a/llvm/include/llvm/CodeGen/CodeGenCommonISel.h
+++ b/llvm/include/llvm/CodeGen/CodeGenCommonISel.h
@@ -218,10 +218,15 @@ findSplitPointForStackProtector(MachineBasicBlock *BB,
/// Evaluates if the specified FP class test is better performed as the inverse
/// (i.e. fewer instructions should be required to lower it). An example is the
/// test "inf|normal|subnormal|zero", which is an inversion of "nan".
+///
/// \param Test The test as specified in 'is_fpclass' intrinsic invocation.
+///
+/// \param UseFCmp The intention is to perform the comparison using
+/// floating-point compare instructions which check for nan.
+///
/// \returns The inverted test, or fcNone, if inversion does not produce a
/// simpler test.
-FPClassTest invertFPClassTestIfSimpler(FPClassTest Test);
+FPClassTest invertFPClassTestIfSimpler(FPClassTest Test, bool UseFCmp);
/// Assuming the instruction \p MI is going to be deleted, attempt to salvage
/// debug users of \p MI by writing the effect of \p MI in a DIExpression.
diff --git a/llvm/lib/CodeGen/CodeGenCommonISel.cpp b/llvm/lib/CodeGen/CodeGenCommonISel.cpp
index fe144d3c182039..d985751e2be0be 100644
--- a/llvm/lib/CodeGen/CodeGenCommonISel.cpp
+++ b/llvm/lib/CodeGen/CodeGenCommonISel.cpp
@@ -173,8 +173,9 @@ llvm::findSplitPointForStackProtector(MachineBasicBlock *BB,
return SplitPoint;
}
-FPClassTest llvm::invertFPClassTestIfSimpler(FPClassTest Test) {
+FPClassTest llvm::invertFPClassTestIfSimpler(FPClassTest Test, bool UseFCmp) {
FPClassTest InvertedTest = ~Test;
+
// Pick the direction with fewer tests
// TODO: Handle more combinations of cases that can be handled together
switch (static_cast<unsigned>(InvertedTest)) {
@@ -200,6 +201,11 @@ FPClassTest llvm::invertFPClassTestIfSimpler(FPClassTest Test) {
case fcSubnormal | fcZero:
case fcSubnormal | fcZero | fcNan:
return InvertedTest;
+ case fcInf | fcNan:
+ // If we're trying to use fcmp, we can take advantage of the nan check
+ // behavior of the compare (but this is more instructions in the integer
+ // expansion).
+ return UseFCmp ? InvertedTest : fcNone;
default:
return fcNone;
}
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 4e796289cff0a1..1e3a0da0f3be5b 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -8672,7 +8672,7 @@ SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,
// Degenerated cases.
if (Test == fcNone)
return DAG.getBoolConstant(false, DL, ResultVT, OperandVT);
- if ((Test & fcAllFlags) == fcAllFlags)
+ if (Test == fcAllFlags)
return DAG.getBoolConstant(true, DL, ResultVT, OperandVT);
// PPC double double is a pair of doubles, of which the higher part determines
@@ -8683,14 +8683,6 @@ SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,
OperandVT = MVT::f64;
}
- // Some checks may be represented as inversion of simpler check, for example
- // "inf|normal|subnormal|zero" => !"nan".
- bool IsInverted = false;
- if (FPClassTest InvertedCheck = invertFPClassTestIfSimpler(Test)) {
- IsInverted = true;
- Test = InvertedCheck;
- }
-
// Floating-point type properties.
EVT ScalarFloatVT = OperandVT.getScalarType();
const Type *FloatTy = ScalarFloatVT.getTypeForEVT(*DAG.getContext());
@@ -8702,9 +8694,16 @@ SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,
if (Flags.hasNoFPExcept() &&
isOperationLegalOrCustom(ISD::SETCC, OperandVT.getScalarType())) {
FPClassTest FPTestMask = Test;
+ bool IsInvertedFP = false;
+
+ if (FPClassTest InvertedFPCheck =
+ invertFPClassTestIfSimpler(FPTestMask, true)) {
+ FPTestMask = InvertedFPCheck;
+ IsInvertedFP = true;
+ }
- ISD::CondCode OrderedCmpOpcode = IsInverted ? ISD::SETUNE : ISD::SETOEQ;
- ISD::CondCode UnorderedCmpOpcode = IsInverted ? ISD::SETONE : ISD::SETUEQ;
+ ISD::CondCode OrderedCmpOpcode = IsInvertedFP ? ISD::SETUNE : ISD::SETOEQ;
+ ISD::CondCode UnorderedCmpOpcode = IsInvertedFP ? ISD::SETONE : ISD::SETUEQ;
// See if we can fold an | fcNan into an unordered compare.
FPClassTest OrderedFPTestMask = FPTestMask & ~fcNan;
@@ -8717,7 +8716,7 @@ SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,
const bool IsOrdered = FPTestMask == OrderedFPTestMask;
if (std::optional<bool> IsCmp0 =
- isFCmpEqualZero(Test, Semantics, DAG.getMachineFunction());
+ isFCmpEqualZero(FPTestMask, Semantics, DAG.getMachineFunction());
IsCmp0 && (isCondCodeLegalOrCustom(
*IsCmp0 ? OrderedCmpOpcode : UnorderedCmpOpcode,
OperandVT.getScalarType().getSimpleVT()))) {
@@ -8729,15 +8728,16 @@ SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,
*IsCmp0 ? OrderedCmpOpcode : UnorderedCmpOpcode);
}
- if (Test == fcNan &&
- isCondCodeLegalOrCustom(IsInverted ? ISD::SETO : ISD::SETUO,
- OperandVT.getScalarType().getSimpleVT())) {
+ if (FPTestMask == fcNan &&
+ isCondCodeLegalOrCustom(IsInvertedFP ? ISD::SETO : ISD::SETUO,
+ OperandVT.getScalarType().getSimpleVT()))
return DAG.getSetCC(DL, ResultVT, Op, Op,
- IsInverted ? ISD::SETO : ISD::SETUO);
- }
+ IsInvertedFP ? ISD::SETO : ISD::SETUO);
- if (Test == fcInf &&
- isCondCodeLegalOrCustom(IsInverted ? ISD::SETUNE : ISD::SETOEQ,
+ bool IsOrderedInf = FPTestMask == fcInf;
+ if ((FPTestMask == fcInf || FPTestMask == (fcInf | fcNan)) &&
+ isCondCodeLegalOrCustom(IsOrderedInf ? OrderedCmpOpcode
+ : UnorderedCmpOpcode,
OperandVT.getScalarType().getSimpleVT()) &&
isOperationLegalOrCustom(ISD::FABS, OperandVT.getScalarType())) {
// isinf(x) --> fabs(x) == inf
@@ -8745,15 +8745,15 @@ SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,
SDValue Inf =
DAG.getConstantFP(APFloat::getInf(Semantics), DL, OperandVT);
return DAG.getSetCC(DL, ResultVT, Abs, Inf,
- IsInverted ? ISD::SETUNE : ISD::SETOEQ);
+ IsOrderedInf ? OrderedCmpOpcode : UnorderedCmpOpcode);
}
if (OrderedFPTestMask == (fcSubnormal | fcZero) && !IsOrdered) {
// TODO: Could handle ordered case, but it produces worse code for
// x86. Maybe handle ordered if fabs is free?
- ISD::CondCode OrderedOp = IsInverted ? ISD::SETUGE : ISD::SETOLT;
- ISD::CondCode UnorderedOp = IsInverted ? ISD::SETOGE : ISD::SETULT;
+ ISD::CondCode OrderedOp = IsInvertedFP ? ISD::SETUGE : ISD::SETOLT;
+ ISD::CondCode UnorderedOp = IsInvertedFP ? ISD::SETOGE : ISD::SETULT;
if (isCondCodeLegalOrCustom(IsOrdered ? OrderedOp : UnorderedOp,
OperandVT.getScalarType().getSimpleVT())) {
@@ -8770,6 +8770,15 @@ SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,
}
}
+ // Some checks may be represented as inversion of simpler check, for example
+ // "inf|normal|subnormal|zero" => !"nan".
+ bool IsInverted = false;
+
+ if (FPClassTest InvertedCheck = invertFPClassTestIfSimpler(Test, false)) {
+ Test = InvertedCheck;
+ IsInverted = true;
+ }
+
// In the general case use integer operations.
unsigned BitSize = OperandVT.getScalarSizeInBits();
EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), BitSize);
diff --git a/llvm/test/CodeGen/X86/is_fpclass.ll b/llvm/test/CodeGen/X86/is_fpclass.ll
index cc4d4c4543a515..4e46abbdb01c34 100644
--- a/llvm/test/CodeGen/X86/is_fpclass.ll
+++ b/llvm/test/CodeGen/X86/is_fpclass.ll
@@ -240,18 +240,22 @@ entry:
define i1 @isfinite_f(float %x) {
; X86-LABEL: isfinite_f:
; X86: # %bb.0: # %entry
-; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
-; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: cmpl $2139095040, %eax # imm = 0x7F800000
-; X86-NEXT: setl %al
+; X86-NEXT: flds {{[0-9]+}}(%esp)
+; X86-NEXT: fabs
+; X86-NEXT: flds {{\.?LCPI[0-9]+_[0-9]+}}
+; X86-NEXT: fxch %st(1)
+; X86-NEXT: fucompp
+; X86-NEXT: fnstsw %ax
+; X86-NEXT: # kill: def $ah killed $ah killed $ax
+; X86-NEXT: sahf
+; X86-NEXT: setne %al
; X86-NEXT: retl
;
; X64-LABEL: isfinite_f:
; X64: # %bb.0: # %entry
-; X64-NEXT: movd %xmm0, %eax
-; X64-NEXT: andl $2147483647, %eax # imm = 0x7FFFFFFF
-; X64-NEXT: cmpl $2139095040, %eax # imm = 0x7F800000
-; X64-NEXT: setl %al
+; X64-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-NEXT: setne %al
; X64-NEXT: retq
entry:
%0 = tail call i1 @llvm.is.fpclass.f32(float %x, i32 504) ; 0x1f8 = "finite"
@@ -1150,31 +1154,23 @@ entry:
define i1 @isfinite_d(double %x) {
; X86-LABEL: isfinite_d:
; X86: # %bb.0: # %entry
-; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
-; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: cmpl $2146435072, %eax # imm = 0x7FF00000
-; X86-NEXT: setl %al
+; X86-NEXT: fldl {{[0-9]+}}(%esp)
+; X86-NEXT: fabs
+; X86-NEXT: flds {{\.?LCPI[0-9]+_[0-9]+}}
+; X86-NEXT: fxch %st(1)
+; X86-NEXT: fucompp
+; X86-NEXT: fnstsw %ax
+; X86-NEXT: # kill: def $ah killed $ah killed $ax
+; X86-NEXT: sahf
+; X86-NEXT: setne %al
; X86-NEXT: retl
;
-; X64-GENERIC-LABEL: isfinite_d:
-; X64-GENERIC: # %bb.0: # %entry
-; X64-GENERIC-NEXT: movq %xmm0, %rax
-; X64-GENERIC-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
-; X64-GENERIC-NEXT: andq %rax, %rcx
-; X64-GENERIC-NEXT: movabsq $9218868437227405312, %rax # imm = 0x7FF0000000000000
-; X64-GENERIC-NEXT: cmpq %rax, %rcx
-; X64-GENERIC-NEXT: setl %al
-; X64-GENERIC-NEXT: retq
-;
-; X64-NDD-LABEL: isfinite_d:
-; X64-NDD: # %bb.0: # %entry
-; X64-NDD-NEXT: movq %xmm0, %rax
-; X64-NDD-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
-; X64-NDD-NEXT: andq %rcx, %rax
-; X64-NDD-NEXT: movabsq $9218868437227405312, %rcx # imm = 0x7FF0000000000000
-; X64-NDD-NEXT: cmpq %rcx, %rax
-; X64-NDD-NEXT: setl %al
-; X64-NDD-NEXT: retq
+; X64-LABEL: isfinite_d:
+; X64: # %bb.0: # %entry
+; X64-NEXT: andpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-NEXT: setne %al
+; X64-NEXT: retq
entry:
%0 = tail call i1 @llvm.is.fpclass.f64(double %x, i32 504) ; 0x1f8 = "finite"
ret i1 %0
@@ -2053,18 +2049,22 @@ entry:
define i1 @not_isinf_or_nan_f(float %x) {
; X86-LABEL: not_isinf_or_nan_f:
; X86: # %bb.0: # %entry
-; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
-; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: cmpl $2139095040, %eax # imm = 0x7F800000
-; X86-NEXT: setl %al
+; X86-NEXT: flds {{[0-9]+}}(%esp)
+; X86-NEXT: fabs
+; X86-NEXT: flds {{\.?LCPI[0-9]+_[0-9]+}}
+; X86-NEXT: fxch %st(1)
+; X86-NEXT: fucompp
+; X86-NEXT: fnstsw %ax
+; X86-NEXT: # kill: def $ah killed $ah killed $ax
+; X86-NEXT: sahf
+; X86-NEXT: setne %al
; X86-NEXT: retl
;
; X64-LABEL: not_isinf_or_nan_f:
; X64: # %bb.0: # %entry
-; X64-NEXT: movd %xmm0, %eax
-; X64-NEXT: andl $2147483647, %eax # imm = 0x7FFFFFFF
-; X64-NEXT: cmpl $2139095040, %eax # imm = 0x7F800000
-; X64-NEXT: setl %al
+; X64-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-NEXT: setne %al
; X64-NEXT: retq
entry:
%0 = tail call i1 @llvm.is.fpclass.f32(float %x, i32 504) ; ~(0x204|0x3) = "~(inf|nan)"
>From 5c3bed8aefde9f2d49182fb1704bcbb2a3c49cac Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Wed, 21 Aug 2024 15:42:24 +0400
Subject: [PATCH 2/3] Check for immediate legality
This avoids the x86_32 regressions, at the expense of
several other cases.
---
.../CodeGen/SelectionDAG/TargetLowering.cpp | 5 +-
llvm/test/CodeGen/AArch64/isinf.ll | 12 +--
llvm/test/CodeGen/PowerPC/fp-classify.ll | 28 +++----
llvm/test/CodeGen/X86/is_fpclass.ll | 78 +++++++++----------
4 files changed, 63 insertions(+), 60 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 1e3a0da0f3be5b..aa022480947a7d 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -8739,7 +8739,10 @@ SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,
isCondCodeLegalOrCustom(IsOrderedInf ? OrderedCmpOpcode
: UnorderedCmpOpcode,
OperandVT.getScalarType().getSimpleVT()) &&
- isOperationLegalOrCustom(ISD::FABS, OperandVT.getScalarType())) {
+ isOperationLegalOrCustom(ISD::FABS, OperandVT.getScalarType()) &&
+ (isOperationLegal(ISD::ConstantFP, OperandVT.getScalarType()) ||
+ (OperandVT.isVector() &&
+ isOperationLegalOrCustom(ISD::BUILD_VECTOR, OperandVT)))) {
// isinf(x) --> fabs(x) == inf
SDValue Abs = DAG.getNode(ISD::FABS, DL, OperandVT, Op);
SDValue Inf =
diff --git a/llvm/test/CodeGen/AArch64/isinf.ll b/llvm/test/CodeGen/AArch64/isinf.ll
index 834417b98743a8..e68539bcf07d9c 100644
--- a/llvm/test/CodeGen/AArch64/isinf.ll
+++ b/llvm/test/CodeGen/AArch64/isinf.ll
@@ -26,10 +26,10 @@ define i32 @replace_isinf_call_f16(half %x) {
define i32 @replace_isinf_call_f32(float %x) {
; CHECK-LABEL: replace_isinf_call_f32:
; CHECK: // %bb.0:
-; CHECK-NEXT: fabs s0, s0
+; CHECK-NEXT: fmov w9, s0
; CHECK-NEXT: mov w8, #2139095040 // =0x7f800000
-; CHECK-NEXT: fmov s1, w8
-; CHECK-NEXT: fcmp s0, s1
+; CHECK-NEXT: and w9, w9, #0x7fffffff
+; CHECK-NEXT: cmp w9, w8
; CHECK-NEXT: cset w0, eq
; CHECK-NEXT: ret
%abs = tail call float @llvm.fabs.f32(float %x)
@@ -42,10 +42,10 @@ define i32 @replace_isinf_call_f32(float %x) {
define i32 @replace_isinf_call_f64(double %x) {
; CHECK-LABEL: replace_isinf_call_f64:
; CHECK: // %bb.0:
-; CHECK-NEXT: fabs d0, d0
+; CHECK-NEXT: fmov x9, d0
; CHECK-NEXT: mov x8, #9218868437227405312 // =0x7ff0000000000000
-; CHECK-NEXT: fmov d1, x8
-; CHECK-NEXT: fcmp d0, d1
+; CHECK-NEXT: and x9, x9, #0x7fffffffffffffff
+; CHECK-NEXT: cmp x9, x8
; CHECK-NEXT: cset w0, eq
; CHECK-NEXT: ret
%abs = tail call double @llvm.fabs.f64(double %x)
diff --git a/llvm/test/CodeGen/PowerPC/fp-classify.ll b/llvm/test/CodeGen/PowerPC/fp-classify.ll
index f527b3c48040e7..dc9853ff2e3014 100644
--- a/llvm/test/CodeGen/PowerPC/fp-classify.ll
+++ b/llvm/test/CodeGen/PowerPC/fp-classify.ll
@@ -7,13 +7,13 @@
define zeroext i1 @abs_isinff(float %x) {
; P8-LABEL: abs_isinff:
; P8: # %bb.0: # %entry
-; P8-NEXT: addis 3, 2, .LCPI0_0 at toc@ha
-; P8-NEXT: xsabsdp 0, 1
-; P8-NEXT: li 4, 1
-; P8-NEXT: lfs 1, .LCPI0_0 at toc@l(3)
-; P8-NEXT: li 3, 0
-; P8-NEXT: fcmpu 0, 0, 1
-; P8-NEXT: iseleq 3, 4, 3
+; P8-NEXT: xscvdpspn 0, 1
+; P8-NEXT: lis 4, 32640
+; P8-NEXT: mffprwz 3, 0
+; P8-NEXT: clrlwi 3, 3, 1
+; P8-NEXT: xor 3, 3, 4
+; P8-NEXT: cntlzw 3, 3
+; P8-NEXT: srwi 3, 3, 5
; P8-NEXT: blr
;
; P9-LABEL: abs_isinff:
@@ -32,13 +32,13 @@ entry:
define zeroext i1 @abs_isinf(double %x) {
; P8-LABEL: abs_isinf:
; P8: # %bb.0: # %entry
-; P8-NEXT: addis 3, 2, .LCPI1_0 at toc@ha
-; P8-NEXT: xsabsdp 0, 1
-; P8-NEXT: li 4, 1
-; P8-NEXT: lfs 1, .LCPI1_0 at toc@l(3)
-; P8-NEXT: li 3, 0
-; P8-NEXT: fcmpu 0, 0, 1
-; P8-NEXT: iseleq 3, 4, 3
+; P8-NEXT: mffprd 3, 1
+; P8-NEXT: li 4, 2047
+; P8-NEXT: rldic 4, 4, 52, 1
+; P8-NEXT: clrldi 3, 3, 1
+; P8-NEXT: xor 3, 3, 4
+; P8-NEXT: cntlzd 3, 3
+; P8-NEXT: rldicl 3, 3, 58, 63
; P8-NEXT: blr
;
; P9-LABEL: abs_isinf:
diff --git a/llvm/test/CodeGen/X86/is_fpclass.ll b/llvm/test/CodeGen/X86/is_fpclass.ll
index 4e46abbdb01c34..cc4d4c4543a515 100644
--- a/llvm/test/CodeGen/X86/is_fpclass.ll
+++ b/llvm/test/CodeGen/X86/is_fpclass.ll
@@ -240,22 +240,18 @@ entry:
define i1 @isfinite_f(float %x) {
; X86-LABEL: isfinite_f:
; X86: # %bb.0: # %entry
-; X86-NEXT: flds {{[0-9]+}}(%esp)
-; X86-NEXT: fabs
-; X86-NEXT: flds {{\.?LCPI[0-9]+_[0-9]+}}
-; X86-NEXT: fxch %st(1)
-; X86-NEXT: fucompp
-; X86-NEXT: fnstsw %ax
-; X86-NEXT: # kill: def $ah killed $ah killed $ax
-; X86-NEXT: sahf
-; X86-NEXT: setne %al
+; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
+; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
+; X86-NEXT: cmpl $2139095040, %eax # imm = 0x7F800000
+; X86-NEXT: setl %al
; X86-NEXT: retl
;
; X64-LABEL: isfinite_f:
; X64: # %bb.0: # %entry
-; X64-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; X64-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; X64-NEXT: setne %al
+; X64-NEXT: movd %xmm0, %eax
+; X64-NEXT: andl $2147483647, %eax # imm = 0x7FFFFFFF
+; X64-NEXT: cmpl $2139095040, %eax # imm = 0x7F800000
+; X64-NEXT: setl %al
; X64-NEXT: retq
entry:
%0 = tail call i1 @llvm.is.fpclass.f32(float %x, i32 504) ; 0x1f8 = "finite"
@@ -1154,23 +1150,31 @@ entry:
define i1 @isfinite_d(double %x) {
; X86-LABEL: isfinite_d:
; X86: # %bb.0: # %entry
-; X86-NEXT: fldl {{[0-9]+}}(%esp)
-; X86-NEXT: fabs
-; X86-NEXT: flds {{\.?LCPI[0-9]+_[0-9]+}}
-; X86-NEXT: fxch %st(1)
-; X86-NEXT: fucompp
-; X86-NEXT: fnstsw %ax
-; X86-NEXT: # kill: def $ah killed $ah killed $ax
-; X86-NEXT: sahf
-; X86-NEXT: setne %al
+; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
+; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
+; X86-NEXT: cmpl $2146435072, %eax # imm = 0x7FF00000
+; X86-NEXT: setl %al
; X86-NEXT: retl
;
-; X64-LABEL: isfinite_d:
-; X64: # %bb.0: # %entry
-; X64-NEXT: andpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; X64-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; X64-NEXT: setne %al
-; X64-NEXT: retq
+; X64-GENERIC-LABEL: isfinite_d:
+; X64-GENERIC: # %bb.0: # %entry
+; X64-GENERIC-NEXT: movq %xmm0, %rax
+; X64-GENERIC-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
+; X64-GENERIC-NEXT: andq %rax, %rcx
+; X64-GENERIC-NEXT: movabsq $9218868437227405312, %rax # imm = 0x7FF0000000000000
+; X64-GENERIC-NEXT: cmpq %rax, %rcx
+; X64-GENERIC-NEXT: setl %al
+; X64-GENERIC-NEXT: retq
+;
+; X64-NDD-LABEL: isfinite_d:
+; X64-NDD: # %bb.0: # %entry
+; X64-NDD-NEXT: movq %xmm0, %rax
+; X64-NDD-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
+; X64-NDD-NEXT: andq %rcx, %rax
+; X64-NDD-NEXT: movabsq $9218868437227405312, %rcx # imm = 0x7FF0000000000000
+; X64-NDD-NEXT: cmpq %rcx, %rax
+; X64-NDD-NEXT: setl %al
+; X64-NDD-NEXT: retq
entry:
%0 = tail call i1 @llvm.is.fpclass.f64(double %x, i32 504) ; 0x1f8 = "finite"
ret i1 %0
@@ -2049,22 +2053,18 @@ entry:
define i1 @not_isinf_or_nan_f(float %x) {
; X86-LABEL: not_isinf_or_nan_f:
; X86: # %bb.0: # %entry
-; X86-NEXT: flds {{[0-9]+}}(%esp)
-; X86-NEXT: fabs
-; X86-NEXT: flds {{\.?LCPI[0-9]+_[0-9]+}}
-; X86-NEXT: fxch %st(1)
-; X86-NEXT: fucompp
-; X86-NEXT: fnstsw %ax
-; X86-NEXT: # kill: def $ah killed $ah killed $ax
-; X86-NEXT: sahf
-; X86-NEXT: setne %al
+; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
+; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
+; X86-NEXT: cmpl $2139095040, %eax # imm = 0x7F800000
+; X86-NEXT: setl %al
; X86-NEXT: retl
;
; X64-LABEL: not_isinf_or_nan_f:
; X64: # %bb.0: # %entry
-; X64-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; X64-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; X64-NEXT: setne %al
+; X64-NEXT: movd %xmm0, %eax
+; X64-NEXT: andl $2147483647, %eax # imm = 0x7FFFFFFF
+; X64-NEXT: cmpl $2139095040, %eax # imm = 0x7F800000
+; X64-NEXT: setl %al
; X64-NEXT: retq
entry:
%0 = tail call i1 @llvm.is.fpclass.f32(float %x, i32 504) ; ~(0x204|0x3) = "~(inf|nan)"
>From 238f697ae0fd38a8a6bfb0c011c416d5a11d71b6 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Thu, 29 Aug 2024 08:17:49 +0400
Subject: [PATCH 3/3] Remove doxygen line
---
llvm/include/llvm/CodeGen/CodeGenCommonISel.h | 1 -
1 file changed, 1 deletion(-)
diff --git a/llvm/include/llvm/CodeGen/CodeGenCommonISel.h b/llvm/include/llvm/CodeGen/CodeGenCommonISel.h
index e4b2e20babc07a..4c22be94507866 100644
--- a/llvm/include/llvm/CodeGen/CodeGenCommonISel.h
+++ b/llvm/include/llvm/CodeGen/CodeGenCommonISel.h
@@ -220,7 +220,6 @@ findSplitPointForStackProtector(MachineBasicBlock *BB,
/// test "inf|normal|subnormal|zero", which is an inversion of "nan".
///
/// \param Test The test as specified in 'is_fpclass' intrinsic invocation.
-///
/// \param UseFCmp The intention is to perform the comparison using
/// floating-point compare instructions which check for nan.
///
More information about the llvm-branch-commits
mailing list