[llvm] SelectionDAG: Support nofpclass with zero/pzero/nzero (PR #137305)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Apr 25 03:01:52 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-x86
Author: YunQiang Su (wzssyqa)
<details>
<summary>Changes</summary>
Add a new argument to isKnownNeverZeroFloat
FPClassTest FPClassArg
so that we can detect any zero (fcZero) or negtive zero (fcNegZero), or postive zero (fcPosZero).
To utilize this change and figure out a test case, we also support it in expandFMINIMUM_FMAXIMUM: compare +0.0 vs -0.0 is not needed if the two arguments are
neither negtive zero
or
neither postive zero
---
Full diff: https://github.com/llvm/llvm-project/pull/137305.diff
5 Files Affected:
- (modified) llvm/include/llvm/CodeGen/SelectionDAG.h (+3-2)
- (modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (+11-1)
- (modified) llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp (+5-1)
- (added) llvm/test/CodeGen/Mips/nofpclass-r2.ll (+117)
- (added) llvm/test/CodeGen/X86/nofpclass.ll (+263)
``````````diff
diff --git a/llvm/include/llvm/CodeGen/SelectionDAG.h b/llvm/include/llvm/CodeGen/SelectionDAG.h
index c183149b0863a..f8774c5209f92 100644
--- a/llvm/include/llvm/CodeGen/SelectionDAG.h
+++ b/llvm/include/llvm/CodeGen/SelectionDAG.h
@@ -2169,8 +2169,9 @@ class SelectionDAG {
}
/// Test whether the given floating point SDValue is known to never be
- /// positive or negative zero.
- bool isKnownNeverZeroFloat(SDValue Op) const;
+ /// positive (FPClass is fcPosZero) or negative zero (FPClass is fcNegZero) or
+ /// any zero (FPClass is fcZero).
+ bool isKnownNeverZeroFloat(SDValue Op, FPClassTest FPClass = fcZero) const;
/// Test whether the given SDValue is known to contain non-zero value(s).
bool isKnownNeverZero(SDValue Op, unsigned Depth = 0) const;
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 9da2ba04f77cb..7a710482de89d 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -5852,10 +5852,20 @@ bool SelectionDAG::isKnownNeverNaN(SDValue Op, const APInt &DemandedElts,
}
}
-bool SelectionDAG::isKnownNeverZeroFloat(SDValue Op) const {
+bool SelectionDAG::isKnownNeverZeroFloat(SDValue Op,
+ FPClassTest FPClassArg) const {
assert(Op.getValueType().isFloatingPoint() &&
"Floating point type expected");
+ FPClassArg &= fcZero;
+ assert(FPClassArg != fcNone && "FPClassArg of isKnownNeverZeroFloat should "
+ "be one of fcZero/fcNegZero/fcPosZero");
+ if (Op->getOpcode() == ISD::AssertNoFPClass) {
+ FPClassTest NoFPClass =
+ static_cast<FPClassTest>(Op.getConstantOperandVal(1));
+ if ((NoFPClass & FPClassArg) == FPClassArg)
+ return true;
+ }
// If the value is a constant, we can obviously see if it is a zero or not.
return ISD::matchUnaryFpPredicate(
Op, [](ConstantFPSDNode *C) { return !C->isZero(); });
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 6930b54ddb14a..19f361f23fbf2 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -8603,7 +8603,11 @@ SDValue TargetLowering::expandFMINIMUM_FMAXIMUM(SDNode *N,
// fminimum/fmaximum requires -0.0 less than +0.0
if (!MinMaxMustRespectOrderedZero && !N->getFlags().hasNoSignedZeros() &&
- !DAG.isKnownNeverZeroFloat(RHS) && !DAG.isKnownNeverZeroFloat(LHS)) {
+ !DAG.isKnownNeverZeroFloat(RHS) && !DAG.isKnownNeverZeroFloat(LHS) &&
+ !(DAG.isKnownNeverZeroFloat(RHS, fcNegZero) &&
+ DAG.isKnownNeverZeroFloat(LHS, fcNegZero)) &&
+ !(DAG.isKnownNeverZeroFloat(RHS, fcPosZero) &&
+ DAG.isKnownNeverZeroFloat(LHS, fcPosZero))) {
SDValue IsZero = DAG.getSetCC(DL, CCVT, MinMax,
DAG.getConstantFP(0.0, DL, VT), ISD::SETOEQ);
SDValue TestZero =
diff --git a/llvm/test/CodeGen/Mips/nofpclass-r2.ll b/llvm/test/CodeGen/Mips/nofpclass-r2.ll
new file mode 100644
index 0000000000000..37fc73038317b
--- /dev/null
+++ b/llvm/test/CodeGen/Mips/nofpclass-r2.ll
@@ -0,0 +1,117 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc --mtriple=mips-linux-gnu < %s | FileCheck %s --check-prefix=MIPS32R2
+
+define float @min(float nofpclass(nan zero) %a, float nofpclass(nan zero) %b) {
+; MIPS32R2-LABEL: min:
+; MIPS32R2: # %bb.0: # %entry
+; MIPS32R2-NEXT: mov.s $f0, $f14
+; MIPS32R2-NEXT: c.olt.s $f12, $f14
+; MIPS32R2-NEXT: jr $ra
+; MIPS32R2-NEXT: movt.s $f0, $f12, $fcc0
+entry:
+ %0 = tail call float @llvm.minimum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @minN(float nofpclass(nan nzero) %a, float nofpclass(nan nzero) %b) {
+; MIPS32R2-LABEL: minN:
+; MIPS32R2: # %bb.0: # %entry
+; MIPS32R2-NEXT: mov.s $f0, $f14
+; MIPS32R2-NEXT: c.olt.s $f12, $f14
+; MIPS32R2-NEXT: jr $ra
+; MIPS32R2-NEXT: movt.s $f0, $f12, $fcc0
+entry:
+ %0 = tail call float @llvm.minimum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @minP(float nofpclass(nan pzero) %a, float nofpclass(nan pzero) %b) {
+; MIPS32R2-LABEL: minP:
+; MIPS32R2: # %bb.0: # %entry
+; MIPS32R2-NEXT: mov.s $f0, $f14
+; MIPS32R2-NEXT: c.olt.s $f12, $f14
+; MIPS32R2-NEXT: jr $ra
+; MIPS32R2-NEXT: movt.s $f0, $f12, $fcc0
+entry:
+ %0 = tail call float @llvm.minimum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @minNP(float nofpclass(nan nzero) %a, float nofpclass(nan pzero) %b) {
+; MIPS32R2-LABEL: minNP:
+; MIPS32R2: # %bb.0: # %entry
+; MIPS32R2-NEXT: c.olt.s $f12, $f14
+; MIPS32R2-NEXT: mov.s $f0, $f14
+; MIPS32R2-NEXT: movt.s $f0, $f12, $fcc0
+; MIPS32R2-NEXT: mfc1 $1, $f12
+; MIPS32R2-NEXT: lui $2, 32768
+; MIPS32R2-NEXT: xor $1, $1, $2
+; MIPS32R2-NEXT: mov.s $f1, $f0
+; MIPS32R2-NEXT: movz.s $f1, $f12, $1
+; MIPS32R2-NEXT: mfc1 $1, $f14
+; MIPS32R2-NEXT: xor $1, $1, $2
+; MIPS32R2-NEXT: movz.s $f1, $f14, $1
+; MIPS32R2-NEXT: mtc1 $zero, $f2
+; MIPS32R2-NEXT: c.eq.s $f0, $f2
+; MIPS32R2-NEXT: jr $ra
+; MIPS32R2-NEXT: movt.s $f0, $f1, $fcc0
+entry:
+ %0 = tail call float @llvm.minimum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @max(float nofpclass(nan zero) %a, float nofpclass(nan zero) %b) {
+; MIPS32R2-LABEL: max:
+; MIPS32R2: # %bb.0: # %entry
+; MIPS32R2-NEXT: mov.s $f0, $f14
+; MIPS32R2-NEXT: c.ule.s $f12, $f14
+; MIPS32R2-NEXT: jr $ra
+; MIPS32R2-NEXT: movf.s $f0, $f12, $fcc0
+entry:
+ %0 = tail call float @llvm.maximum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @maxN(float nofpclass(nan nzero) %a, float nofpclass(nan nzero) %b) {
+; MIPS32R2-LABEL: maxN:
+; MIPS32R2: # %bb.0: # %entry
+; MIPS32R2-NEXT: mov.s $f0, $f14
+; MIPS32R2-NEXT: c.ule.s $f12, $f14
+; MIPS32R2-NEXT: jr $ra
+; MIPS32R2-NEXT: movf.s $f0, $f12, $fcc0
+entry:
+ %0 = tail call float @llvm.maximum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @maxP(float nofpclass(nan pzero) %a, float nofpclass(nan pzero) %b) {
+; MIPS32R2-LABEL: maxP:
+; MIPS32R2: # %bb.0: # %entry
+; MIPS32R2-NEXT: mov.s $f0, $f14
+; MIPS32R2-NEXT: c.ule.s $f12, $f14
+; MIPS32R2-NEXT: jr $ra
+; MIPS32R2-NEXT: movf.s $f0, $f12, $fcc0
+entry:
+ %0 = tail call float @llvm.maximum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @maxNP(float nofpclass(nan nzero) %a, float nofpclass(nan pzero) %b) {
+; MIPS32R2-LABEL: maxNP:
+; MIPS32R2: # %bb.0: # %entry
+; MIPS32R2-NEXT: c.ule.s $f12, $f14
+; MIPS32R2-NEXT: mov.s $f0, $f14
+; MIPS32R2-NEXT: movf.s $f0, $f12, $fcc0
+; MIPS32R2-NEXT: mfc1 $1, $f12
+; MIPS32R2-NEXT: mov.s $f1, $f0
+; MIPS32R2-NEXT: movz.s $f1, $f12, $1
+; MIPS32R2-NEXT: mfc1 $1, $f14
+; MIPS32R2-NEXT: movz.s $f1, $f14, $1
+; MIPS32R2-NEXT: mtc1 $zero, $f2
+; MIPS32R2-NEXT: c.eq.s $f0, $f2
+; MIPS32R2-NEXT: jr $ra
+; MIPS32R2-NEXT: movt.s $f0, $f1, $fcc0
+entry:
+ %0 = tail call float @llvm.maximum.f32(float %a, float %b)
+ ret float %0
+}
diff --git a/llvm/test/CodeGen/X86/nofpclass.ll b/llvm/test/CodeGen/X86/nofpclass.ll
new file mode 100644
index 0000000000000..be97d486df40c
--- /dev/null
+++ b/llvm/test/CodeGen/X86/nofpclass.ll
@@ -0,0 +1,263 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc --mtriple=i686-linux-gnu < %s | FileCheck %s --check-prefix=I686
+
+define float @min(float nofpclass(nan zero) %a, float nofpclass(nan zero) %b) {
+; I686-LABEL: min:
+; I686: # %bb.0: # %entry
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fucom %st(1)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: ja .LBB0_2
+; I686-NEXT: # %bb.1: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: fldz
+; I686-NEXT: .LBB0_2: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: retl
+entry:
+ %0 = tail call float @llvm.minimum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @minN(float nofpclass(nan nzero) %a, float nofpclass(nan nzero) %b) {
+; I686-LABEL: minN:
+; I686: # %bb.0: # %entry
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fucom %st(1)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: ja .LBB1_2
+; I686-NEXT: # %bb.1: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: fldz
+; I686-NEXT: .LBB1_2: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: retl
+entry:
+ %0 = tail call float @llvm.minimum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @minP(float nofpclass(nan pzero) %a, float nofpclass(nan pzero) %b) {
+; I686-LABEL: minP:
+; I686: # %bb.0: # %entry
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fucom %st(1)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: ja .LBB2_2
+; I686-NEXT: # %bb.1: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: fldz
+; I686-NEXT: .LBB2_2: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: retl
+entry:
+ %0 = tail call float @llvm.minimum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @minNP(float nofpclass(nan nzero) %a, float nofpclass(nan pzero) %b) {
+; I686-LABEL: minNP:
+; I686: # %bb.0: # %entry
+; I686-NEXT: subl $8, %esp
+; I686-NEXT: .cfi_def_cfa_offset 12
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fsts {{[0-9]+}}(%esp)
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fsts (%esp)
+; I686-NEXT: fucom %st(1)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: fld %st(1)
+; I686-NEXT: ja .LBB3_2
+; I686-NEXT: # %bb.1: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: fld %st(0)
+; I686-NEXT: .LBB3_2: # %entry
+; I686-NEXT: xorl %eax, %eax
+; I686-NEXT: cmpl {{[0-9]+}}(%esp), %eax
+; I686-NEXT: jo .LBB3_4
+; I686-NEXT: # %bb.3: # %entry
+; I686-NEXT: fstp %st(2)
+; I686-NEXT: fld %st(1)
+; I686-NEXT: fxch %st(2)
+; I686-NEXT: .LBB3_4: # %entry
+; I686-NEXT: cmpl (%esp), %eax
+; I686-NEXT: jo .LBB3_6
+; I686-NEXT: # %bb.5: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: fldz
+; I686-NEXT: fxch %st(2)
+; I686-NEXT: fxch %st(1)
+; I686-NEXT: .LBB3_6: # %entry
+; I686-NEXT: fstp %st(2)
+; I686-NEXT: fldz
+; I686-NEXT: fxch %st(2)
+; I686-NEXT: fucom %st(2)
+; I686-NEXT: fstp %st(2)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: jne .LBB3_7
+; I686-NEXT: # %bb.8: # %entry
+; I686-NEXT: jp .LBB3_11
+; I686-NEXT: # %bb.9: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: jmp .LBB3_10
+; I686-NEXT: .LBB3_7:
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: .LBB3_10: # %entry
+; I686-NEXT: fldz
+; I686-NEXT: .LBB3_11: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: addl $8, %esp
+; I686-NEXT: .cfi_def_cfa_offset 4
+; I686-NEXT: retl
+entry:
+ %0 = tail call float @llvm.minimum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @max(float nofpclass(nan zero) %a, float nofpclass(nan zero) %b) {
+; I686-LABEL: max:
+; I686: # %bb.0: # %entry
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fucom %st(1)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: ja .LBB4_2
+; I686-NEXT: # %bb.1: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: fldz
+; I686-NEXT: fxch %st(1)
+; I686-NEXT: .LBB4_2: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: retl
+entry:
+ %0 = tail call float @llvm.maximum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @maxN(float nofpclass(nan nzero) %a, float nofpclass(nan nzero) %b) {
+; I686-LABEL: maxN:
+; I686: # %bb.0: # %entry
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fucom %st(1)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: ja .LBB5_2
+; I686-NEXT: # %bb.1: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: fldz
+; I686-NEXT: fxch %st(1)
+; I686-NEXT: .LBB5_2: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: retl
+entry:
+ %0 = tail call float @llvm.maximum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @maxP(float nofpclass(nan pzero) %a, float nofpclass(nan pzero) %b) {
+; I686-LABEL: maxP:
+; I686: # %bb.0: # %entry
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fucom %st(1)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: ja .LBB6_2
+; I686-NEXT: # %bb.1: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: fldz
+; I686-NEXT: fxch %st(1)
+; I686-NEXT: .LBB6_2: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: retl
+entry:
+ %0 = tail call float @llvm.maximum.f32(float %a, float %b)
+ ret float %0
+}
+
+define float @maxNP(float nofpclass(nan nzero) %a, float nofpclass(nan pzero) %b) {
+; I686-LABEL: maxNP:
+; I686: # %bb.0: # %entry
+; I686-NEXT: subl $8, %esp
+; I686-NEXT: .cfi_def_cfa_offset 12
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fsts {{[0-9]+}}(%esp)
+; I686-NEXT: flds {{[0-9]+}}(%esp)
+; I686-NEXT: fsts (%esp)
+; I686-NEXT: fxch %st(1)
+; I686-NEXT: fucom %st(1)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: fld %st(0)
+; I686-NEXT: jbe .LBB7_1
+; I686-NEXT: # %bb.2: # %entry
+; I686-NEXT: cmpl $0, {{[0-9]+}}(%esp)
+; I686-NEXT: jne .LBB7_3
+; I686-NEXT: .LBB7_4: # %entry
+; I686-NEXT: cmpl $0, (%esp)
+; I686-NEXT: je .LBB7_6
+; I686-NEXT: .LBB7_5: # %entry
+; I686-NEXT: fstp %st(2)
+; I686-NEXT: fldz
+; I686-NEXT: fxch %st(1)
+; I686-NEXT: fxch %st(2)
+; I686-NEXT: .LBB7_6: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: fldz
+; I686-NEXT: fxch %st(1)
+; I686-NEXT: fucom %st(1)
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: fnstsw %ax
+; I686-NEXT: # kill: def $ah killed $ah killed $ax
+; I686-NEXT: sahf
+; I686-NEXT: jne .LBB7_7
+; I686-NEXT: # %bb.8: # %entry
+; I686-NEXT: jp .LBB7_11
+; I686-NEXT: # %bb.9: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: jmp .LBB7_10
+; I686-NEXT: .LBB7_1: # %entry
+; I686-NEXT: fstp %st(0)
+; I686-NEXT: fld %st(1)
+; I686-NEXT: cmpl $0, {{[0-9]+}}(%esp)
+; I686-NEXT: je .LBB7_4
+; I686-NEXT: .LBB7_3: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: fld %st(0)
+; I686-NEXT: fxch %st(1)
+; I686-NEXT: cmpl $0, (%esp)
+; I686-NEXT: jne .LBB7_5
+; I686-NEXT: jmp .LBB7_6
+; I686-NEXT: .LBB7_7:
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: .LBB7_10: # %entry
+; I686-NEXT: fldz
+; I686-NEXT: fxch %st(1)
+; I686-NEXT: .LBB7_11: # %entry
+; I686-NEXT: fstp %st(1)
+; I686-NEXT: addl $8, %esp
+; I686-NEXT: .cfi_def_cfa_offset 4
+; I686-NEXT: retl
+entry:
+ %0 = tail call float @llvm.maximum.f32(float %a, float %b)
+ ret float %0
+}
``````````
</details>
https://github.com/llvm/llvm-project/pull/137305
More information about the llvm-commits
mailing list