[Attributor][FIX] Replace AANoFPClass MBEC propagation #91030
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
jdoerfert
force-pushed
the
attributor-fix-fp
branch
from
54d9c78
to
28356b7
Compare
arsenm
reviewed
May 4, 2024
Comment on lines
+10340
to
10347
| if (!CB->isArgOperand(U)) | ||
| return false; |
There was a problem hiding this comment.
An FP type / vector couldn't be the arg operand so is this condition even possible?
There was a problem hiding this comment.
Can be a operand bundle entry, I believe.
| @@ -27,8 +27,8 @@ define { float, i32 } @ret_frexp_f32_nonan(float nofpclass(nan) %arg0) { | |||
|
|
|||
| define float @ret_frexp_f32_0_nonan(float nofpclass(nan) %arg0) { | |||
| ; CHECK-LABEL: define nofpclass(nan sub) float @ret_frexp_f32_0_nonan | |||
| ; CHECK-SAME: (float nofpclass(nan sub) [[ARG0:%.*]]) #[[ATTR1]] { | |||
| ; CHECK-NEXT: [[CALL:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[ARG0]]) #[[ATTR7]] | |||
| ; CHECK-SAME: (float nofpclass(nan) [[ARG0:%.*]]) #[[ATTR1]] { | |||
There was a problem hiding this comment.
I think we still need the special case for frexp, this should still imply nofpclass(sub)
There was a problem hiding this comment.
What is the rule then?
This would work already if frexp had the proper argument attribute.
There was a problem hiding this comment.
You cannot apply the attribute to the return value because it's a struct, not an FP/FP vector. You have to specially recognize the extractvalue of the first element of the struct
There was a problem hiding this comment.
I don't get it. Extractvalue does not imply anything.
What you said first, or what I understood, is that frexp implies that the argument is nofpclass(sub). Is that true? If so, I can implement that.
There was a problem hiding this comment.
The floating point component of the return value of frexp cannot be a denormal. The input could be anything. Because it returns a struct, anything you want to say about the frexp result has to be applied to the extractvalue user of frexp, not the frexp directly
There was a problem hiding this comment.
The test changed such that the argument of the function, the argument to frexp, is not assumed non-sub. What is the problem here?
There was a problem hiding this comment.
What is the problem here?
That I'm reading the argument attribute as the return
|
@llvm/pr-subscribers-llvm-transforms Author: Johannes Doerfert (jdoerfert) ChangesThe old must-be-executed-context (MBEC) propagation did propagate Fixes: #78507 Patch is 720.89 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/91030.diff 35 Files Affected:
diff --git a/llvm/include/llvm/Transforms/IPO/Attributor.h b/llvm/include/llvm/Transforms/IPO/Attributor.h
index 30c51250af61c..d3d3a9c43c84f 100644
--- a/llvm/include/llvm/Transforms/IPO/Attributor.h
+++ b/llvm/include/llvm/Transforms/IPO/Attributor.h
@@ -5418,10 +5418,14 @@ struct AANoFPClass
return false;
}
- /// Return true if we assume that the underlying value is nofpclass.
+ /// Return the underlying assumed nofpclass.
FPClassTest getAssumedNoFPClass() const {
return static_cast<FPClassTest>(getAssumed());
}
+ /// Return the underlying known nofpclass.
+ FPClassTest getKnownNoFPClass() const {
+ return static_cast<FPClassTest>(getKnown());
+ }
/// Create an abstract attribute view for the position \p IRP.
static AANoFPClass &createForPosition(const IRPosition &IRP, Attributor &A);
diff --git a/llvm/lib/Transforms/IPO/AttributorAttributes.cpp b/llvm/lib/Transforms/IPO/AttributorAttributes.cpp
index 41b66aafe7d34..5ddcf9ec14514 100644
--- a/llvm/lib/Transforms/IPO/AttributorAttributes.cpp
+++ b/llvm/lib/Transforms/IPO/AttributorAttributes.cpp
@@ -10332,48 +10332,25 @@ struct AANoFPClassImpl : AANoFPClass {
/// See followUsesInMBEC
bool followUseInMBEC(Attributor &A, const Use *U, const Instruction *I,
AANoFPClass::StateType &State) {
- const Value *UseV = U->get();
- const DominatorTree *DT = nullptr;
- AssumptionCache *AC = nullptr;
- const TargetLibraryInfo *TLI = nullptr;
- InformationCache &InfoCache = A.getInfoCache();
-
- if (Function *F = getAnchorScope()) {
- DT = InfoCache.getAnalysisResultForFunction<DominatorTreeAnalysis>(*F);
- AC = InfoCache.getAnalysisResultForFunction<AssumptionAnalysis>(*F);
- TLI = InfoCache.getTargetLibraryInfoForFunction(*F);
- }
-
- const DataLayout &DL = A.getDataLayout();
-
- KnownFPClass KnownFPClass =
- computeKnownFPClass(UseV, DL,
- /*InterestedClasses=*/fcAllFlags,
- /*Depth=*/0, TLI, AC, I, DT);
- State.addKnownBits(~KnownFPClass.KnownFPClasses);
-
- if (auto *CI = dyn_cast<CallInst>(UseV)) {
- // Special case FP intrinsic with struct return type.
- switch (CI->getIntrinsicID()) {
- case Intrinsic::frexp:
- return true;
- case Intrinsic::not_intrinsic:
- // TODO: Could recognize math libcalls
- return false;
- default:
- break;
- }
- }
+ // TODO: Determine what instructions can be looked through.
+ auto *CB = dyn_cast<CallBase>(I);
+ if (!CB)
+ return false;
- if (!UseV->getType()->isFPOrFPVectorTy())
+ if (!CB->isArgOperand(U))
return false;
- return !isa<LoadInst, AtomicRMWInst>(UseV);
+
+ unsigned ArgNo = CB->getArgOperandNo(U);
+ IRPosition IRP = IRPosition::callsite_argument(*CB, ArgNo);
+ if (auto *NoFPAA = A.getAAFor<AANoFPClass>(*this, IRP, DepClassTy::NONE))
+ State.addKnownBits(NoFPAA->getState().getKnown());
+ return false;
}
const std::string getAsStr(Attributor *A) const override {
std::string Result = "nofpclass";
raw_string_ostream OS(Result);
- OS << getAssumedNoFPClass();
+ OS << getKnownNoFPClass() << '/' << getAssumedNoFPClass();
return Result;
}
diff --git a/llvm/test/Transforms/Attributor/nofpclass-arithmetic-fence.ll b/llvm/test/Transforms/Attributor/nofpclass-arithmetic-fence.ll
index 2b2418cdfde91..c8817ade42098 100644
--- a/llvm/test/Transforms/Attributor/nofpclass-arithmetic-fence.ll
+++ b/llvm/test/Transforms/Attributor/nofpclass-arithmetic-fence.ll
@@ -16,7 +16,7 @@ define float @ret_arithmetic.fence(float %arg0) {
define float @ret_arithmetic.fence_noinf(float nofpclass(inf) %arg0) {
; CHECK-LABEL: define nofpclass(inf) float @ret_arithmetic.fence_noinf
; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf) float @llvm.arithmetic.fence.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
@@ -26,7 +26,7 @@ define float @ret_arithmetic.fence_noinf(float nofpclass(inf) %arg0) {
define float @ret_arithmetic.fence_nopinf(float nofpclass(pinf) %arg0) {
; CHECK-LABEL: define nofpclass(pinf) float @ret_arithmetic.fence_nopinf
; CHECK-SAME: (float nofpclass(pinf) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(pinf) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(pinf) float @llvm.arithmetic.fence.f32(float nofpclass(pinf) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
@@ -36,7 +36,7 @@ define float @ret_arithmetic.fence_nopinf(float nofpclass(pinf) %arg0) {
define float @ret_arithmetic.fence_noninf(float nofpclass(ninf) %arg0) {
; CHECK-LABEL: define nofpclass(ninf) float @ret_arithmetic.fence_noninf
; CHECK-SAME: (float nofpclass(ninf) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(ninf) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(ninf) float @llvm.arithmetic.fence.f32(float nofpclass(ninf) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
@@ -46,7 +46,7 @@ define float @ret_arithmetic.fence_noninf(float nofpclass(ninf) %arg0) {
define float @ret_arithmetic.fence_nonan(float nofpclass(nan) %arg0) {
; CHECK-LABEL: define nofpclass(nan) float @ret_arithmetic.fence_nonan
; CHECK-SAME: (float nofpclass(nan) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(nan) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(nan) float @llvm.arithmetic.fence.f32(float nofpclass(nan) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
@@ -56,7 +56,7 @@ define float @ret_arithmetic.fence_nonan(float nofpclass(nan) %arg0) {
define float @ret_arithmetic.fence_noqnan(float nofpclass(qnan) %arg0) {
; CHECK-LABEL: define nofpclass(qnan) float @ret_arithmetic.fence_noqnan
; CHECK-SAME: (float nofpclass(qnan) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(qnan) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(qnan) float @llvm.arithmetic.fence.f32(float nofpclass(qnan) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
@@ -66,7 +66,7 @@ define float @ret_arithmetic.fence_noqnan(float nofpclass(qnan) %arg0) {
define float @ret_arithmetic.fence_nosnan(float nofpclass(snan) %arg0) {
; CHECK-LABEL: define nofpclass(snan) float @ret_arithmetic.fence_nosnan
; CHECK-SAME: (float nofpclass(snan) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan) float @llvm.arithmetic.fence.f32(float nofpclass(snan) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
@@ -76,7 +76,7 @@ define float @ret_arithmetic.fence_nosnan(float nofpclass(snan) %arg0) {
define float @ret_arithmetic.fence_nozero(float nofpclass(zero) %arg0) {
; CHECK-LABEL: define nofpclass(zero) float @ret_arithmetic.fence_nozero
; CHECK-SAME: (float nofpclass(zero) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(zero) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(zero) float @llvm.arithmetic.fence.f32(float nofpclass(zero) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
@@ -86,7 +86,7 @@ define float @ret_arithmetic.fence_nozero(float nofpclass(zero) %arg0) {
define float @ret_arithmetic.fence_nopzero(float nofpclass(pzero) %arg0) {
; CHECK-LABEL: define nofpclass(pzero) float @ret_arithmetic.fence_nopzero
; CHECK-SAME: (float nofpclass(pzero) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(pzero) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(pzero) float @llvm.arithmetic.fence.f32(float nofpclass(pzero) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
@@ -96,7 +96,7 @@ define float @ret_arithmetic.fence_nopzero(float nofpclass(pzero) %arg0) {
define float @ret_arithmetic.fence_nonzero(float nofpclass(nzero) %arg0) {
; CHECK-LABEL: define nofpclass(nzero) float @ret_arithmetic.fence_nonzero
; CHECK-SAME: (float nofpclass(nzero) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(nzero) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(nzero) float @llvm.arithmetic.fence.f32(float nofpclass(nzero) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
@@ -106,7 +106,7 @@ define float @ret_arithmetic.fence_nonzero(float nofpclass(nzero) %arg0) {
define float @ret_arithmetic.fence_nonorm(float nofpclass(norm) %arg0) {
; CHECK-LABEL: define nofpclass(norm) float @ret_arithmetic.fence_nonorm
; CHECK-SAME: (float nofpclass(norm) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(norm) float @llvm.arithmetic.fence.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(norm) float @llvm.arithmetic.fence.f32(float nofpclass(norm) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.arithmetic.fence.f32(float %arg0)
diff --git a/llvm/test/Transforms/Attributor/nofpclass-canonicalize.ll b/llvm/test/Transforms/Attributor/nofpclass-canonicalize.ll
index 4d3275c3ab1a2..35f8456828ca5 100644
--- a/llvm/test/Transforms/Attributor/nofpclass-canonicalize.ll
+++ b/llvm/test/Transforms/Attributor/nofpclass-canonicalize.ll
@@ -5,7 +5,7 @@ declare float @llvm.canonicalize.f32(float)
define float @ret_canonicalize(float %arg0) {
; CHECK-LABEL: define nofpclass(snan) float @ret_canonicalize
-; CHECK-SAME: (float nofpclass(snan) [[ARG0:%.*]]) #[[ATTR1:[0-9]+]] {
+; CHECK-SAME: (float [[ARG0:%.*]]) #[[ATTR1:[0-9]+]] {
; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12:[0-9]+]]
; CHECK-NEXT: ret float [[CALL]]
;
@@ -15,8 +15,8 @@ define float @ret_canonicalize(float %arg0) {
define float @ret_canonicalize_noinf(float nofpclass(inf) %arg0) {
; CHECK-LABEL: define nofpclass(snan inf) float @ret_canonicalize_noinf
-; CHECK-SAME: (float nofpclass(snan inf) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR1]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf) float @llvm.canonicalize.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR12]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.canonicalize.f32(float %arg0)
@@ -25,8 +25,8 @@ define float @ret_canonicalize_noinf(float nofpclass(inf) %arg0) {
define float @ret_canonicalize_dynamic_denormal(float nofpclass(inf) %arg0) "denormal-fp-math"="dynamic,dynamic" {
; CHECK-LABEL: define nofpclass(snan inf) float @ret_canonicalize_dynamic_denormal
-; CHECK-SAME: (float nofpclass(snan inf) [[ARG0:%.*]]) #[[ATTR2:[0-9]+]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR2:[0-9]+]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf) float @llvm.canonicalize.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR12]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.canonicalize.f32(float %arg0)
@@ -35,8 +35,8 @@ define float @ret_canonicalize_dynamic_denormal(float nofpclass(inf) %arg0) "den
define float @ret_canonicalize_daz_denormal(float nofpclass(inf) %arg0) "denormal-fp-math"="dynamic,preserve-sign" {
; CHECK-LABEL: define nofpclass(snan inf sub) float @ret_canonicalize_daz_denormal
-; CHECK-SAME: (float nofpclass(snan inf sub) [[ARG0:%.*]]) #[[ATTR3:[0-9]+]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf sub) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR3:[0-9]+]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf sub) float @llvm.canonicalize.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR12]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.canonicalize.f32(float %arg0)
@@ -45,8 +45,8 @@ define float @ret_canonicalize_daz_denormal(float nofpclass(inf) %arg0) "denorma
define float @ret_canonicalize_dapz_denormal(float nofpclass(inf) %arg0) "denormal-fp-math"="dynamic,positive-zero" {
; CHECK-LABEL: define nofpclass(snan inf nzero sub) float @ret_canonicalize_dapz_denormal
-; CHECK-SAME: (float nofpclass(snan inf nzero sub) [[ARG0:%.*]]) #[[ATTR4:[0-9]+]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf nzero sub) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR4:[0-9]+]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf nzero sub) float @llvm.canonicalize.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR12]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.canonicalize.f32(float %arg0)
@@ -55,8 +55,8 @@ define float @ret_canonicalize_dapz_denormal(float nofpclass(inf) %arg0) "denorm
define float @ret_canonicalize_ftpz_dapz_denormal(float nofpclass(inf) %arg0) "denormal-fp-math"="positive-zero,preserve-sign" {
; CHECK-LABEL: define nofpclass(snan inf sub) float @ret_canonicalize_ftpz_dapz_denormal
-; CHECK-SAME: (float nofpclass(snan inf sub) [[ARG0:%.*]]) #[[ATTR5:[0-9]+]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf sub) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR5:[0-9]+]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf sub) float @llvm.canonicalize.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR12]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.canonicalize.f32(float %arg0)
@@ -65,8 +65,8 @@ define float @ret_canonicalize_ftpz_dapz_denormal(float nofpclass(inf) %arg0) "d
define float @ret_canonicalize_ftpz_ieee_denormal(float nofpclass(inf) %arg0) "denormal-fp-math"="positive-zero,ieee" {
; CHECK-LABEL: define nofpclass(snan inf nzero sub) float @ret_canonicalize_ftpz_ieee_denormal
-; CHECK-SAME: (float nofpclass(snan inf nzero sub) [[ARG0:%.*]]) #[[ATTR6:[0-9]+]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf nzero sub) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR6:[0-9]+]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf nzero sub) float @llvm.canonicalize.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR12]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.canonicalize.f32(float %arg0)
@@ -75,8 +75,8 @@ define float @ret_canonicalize_ftpz_ieee_denormal(float nofpclass(inf) %arg0) "d
define float @ret_canonicalize_ftpz_dynamic_denormal(float nofpclass(inf) %arg0) "denormal-fp-math"="positive-zero,dynamic" {
; CHECK-LABEL: define nofpclass(snan inf sub) float @ret_canonicalize_ftpz_dynamic_denormal
-; CHECK-SAME: (float nofpclass(snan inf sub) [[ARG0:%.*]]) #[[ATTR7:[0-9]+]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf sub) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR7:[0-9]+]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf sub) float @llvm.canonicalize.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR12]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.canonicalize.f32(float %arg0)
@@ -85,8 +85,8 @@ define float @ret_canonicalize_ftpz_dynamic_denormal(float nofpclass(inf) %arg0)
define float @ret_canonicalize_ftz_denormal(float nofpclass(inf) %arg0) "denormal-fp-math"="preserve-sign,dynamic" {
; CHECK-LABEL: define nofpclass(snan inf sub) float @ret_canonicalize_ftz_denormal
-; CHECK-SAME: (float nofpclass(snan inf sub) [[ARG0:%.*]]) #[[ATTR8:[0-9]+]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf sub) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR8:[0-9]+]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf sub) float @llvm.canonicalize.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR12]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.canonicalize.f32(float %arg0)
@@ -95,8 +95,8 @@ define float @ret_canonicalize_ftz_denormal(float nofpclass(inf) %arg0) "denorma
define float @ret_canonicalize_ieee_denormal(float nofpclass(inf) %arg0) "denormal-fp-math"="ieee,ieee" {
; CHECK-LABEL: define nofpclass(snan inf) float @ret_canonicalize_ieee_denormal
-; CHECK-SAME: (float nofpclass(snan inf) [[ARG0:%.*]]) #[[ATTR9:[0-9]+]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf) float @llvm.canonicalize.f32(float [[ARG0]]) #[[ATTR12]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR9:[0-9]+]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf) float @llvm.canonicalize.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR12]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.canonicalize.f32(float %arg0)
diff --git a/llvm/test/Transforms/Attributor/nofpclass-ceil.ll b/llvm/test/Transforms/Attributor/nofpclass-ceil.ll
index 5f639b2e26917..27a56f2ca9e53 100644
--- a/llvm/test/Transforms/Attributor/nofpclass-ceil.ll
+++ b/llvm/test/Transforms/Attributor/nofpclass-ceil.ll
@@ -6,7 +6,7 @@ declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128)
define float @ret_ceil(float %arg0) {
; CHECK-LABEL: define nofpclass(sub) float @ret_ceil
-; CHECK-SAME: (float nofpclass(sub) [[ARG0:%.*]]) #[[ATTR1:[0-9]+]] {
+; CHECK-SAME: (float [[ARG0:%.*]]) #[[ATTR1:[0-9]+]] {
; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(sub) float @llvm.ceil.f32(float [[ARG0]]) #[[ATTR2:[0-9]+]]
; CHECK-NEXT: ret float [[CALL]]
;
@@ -16,8 +16,8 @@ define float @ret_ceil(float %arg0) {
define float @ret_ceil_noinf(float nofpclass(inf) %arg0) {
; CHECK-LABEL: define nofpclass(inf sub) float @ret_ceil_noinf
-; CHECK-SAME: (float nofpclass(inf sub) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf sub) float @llvm.ceil.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR1]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf sub) float @llvm.ceil.f32(float nofpclass(inf) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.ceil.f32(float %arg0)
@@ -26,8 +26,8 @@ define float @ret_ceil_noinf(float nofpclass(inf) %arg0) {
define float @ret_ceil_nopinf(float nofpclass(pinf) %arg0) {
; CHECK-LABEL: define nofpclass(pinf sub) float @ret_ceil_nopinf
-; CHECK-SAME: (float nofpclass(pinf sub) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(pinf sub) float @llvm.ceil.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-SAME: (float nofpclass(pinf) [[ARG0:%.*]]) #[[ATTR1]] {
+; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(pinf sub) float @llvm.ceil.f32(float nofpclass(pinf) [[ARG0]]) #[[ATTR2]]
; CHECK-NEXT: ret float [[CALL]]
;
%call = call float @llvm.ceil.f32(float %arg0)
@@ -36,8 +36,8 @@ define float @ret_ceil_nopinf(float nofpclass(pinf) %arg0) {
define float @ret_ceil_noninf(float nofpclass(ninf) %arg0) {
; CHECK-LABEL: define nofpclass(ninf sub) float @ret_ceil_noninf
-; CHECK-SAME: (float nofpclass(ninf sub) [[ARG0:%.*]]) #[[ATTR1]] {
-; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(ninf sub) float @llvm.ceil.f32(float [[ARG0]]) #[[ATTR2]]
+; CHECK-SAME: (float nofpclass(ninf) [[ARG0:%.*]]) #[[ATTR1]] {
+; CHECK-NEXT: [[CALL:%.*]] = ...
[truncated]
|
jdoerfert
force-pushed
the
attributor-fix-fp
branch
from
4aea081
to
ae7bd24
Compare
arsenm
reviewed
May 31, 2024
| @@ -27,8 +27,8 @@ define { float, i32 } @ret_frexp_f32_nonan(float nofpclass(nan) %arg0) { | |||
|
|
|||
| define float @ret_frexp_f32_0_nonan(float nofpclass(nan) %arg0) { | |||
| ; CHECK-LABEL: define nofpclass(nan sub) float @ret_frexp_f32_0_nonan | |||
| ; CHECK-SAME: (float nofpclass(nan sub) [[ARG0:%.*]]) #[[ATTR1]] { | |||
| ; CHECK-NEXT: [[CALL:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[ARG0]]) #[[ATTR7]] | |||
| ; CHECK-SAME: (float nofpclass(nan) [[ARG0:%.*]]) #[[ATTR1]] { | |||
There was a problem hiding this comment.
What is the problem here?
That I'm reading the argument attribute as the return
| br i1 %1, label %2, label %21 | ||
|
|
||
| 2: ; preds = %entry | ||
| %3 = fmul double %0, 0x3FE45F306DC9C883 |
There was a problem hiding this comment.
Use named values, and presumably this can be reduced a lot?
Comment on lines
340
to
343
| !opencl.ocl.version = !{!0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0} | ||
| !llvm.ident = !{!1, !2, !1, !2, !1, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2, !1, !2} | ||
| !llvm.module.flags = !{!3, !4, !5, !6, !7, !8, !9} | ||
| !omp_offload.info = !{!10, !11, !12} |
| } | ||
|
|
||
| attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) } | ||
| attributes #1 = { alwaysinline mustprogress nofree norecurse nosync nounwind willreturn memory(none) "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="gfx1030" "target-features"="+16-bit-insts,+ci-insts,+dl-insts,+dot1-insts,+dot10-insts,+dot2-insts,+dot5-insts,+dot6-insts,+dot7-insts,+dpp,+gfx10-3-insts,+gfx10-insts,+gfx8-insts,+gfx9-insts,+s-memrealtime,+s-memtime-inst,+wavefrontsize32" } |
There was a problem hiding this comment.
Drop all the target attributes
The old must-be-executed-context (MBEC) propagation did propagate through calls even if that was not allowed. We now only propagate from call site arguments. If there are calls/intrinsics that allows propagation, we need to add them explicitly. Fixes: llvm#78507
Co-authored-by: Matt Arsenault <arsenm2@gmail.com>
Co-authored-by: Matt Arsenault <arsenm2@gmail.com>
jdoerfert
force-pushed
the
attributor-fix-fp
branch
from
209704c
to
0c12114
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The old must-be-executed-context (MBEC) propagation did propagate
through calls even if that was not allowed. We now only propagate from
call site arguments. If there are calls/intrinsics that allows
propagation, we need to add them explicitly.
Fixes: #78507