ValueTracking: Teach computeKnownFPClass that multiply can avoid denormals #171730
+29
−11
Conversation
This file contains hidden or 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
This was referenced Dec 10, 2025
Contributor
Author
This was referenced Dec 10, 2025
arsenm
added
floating-point
Graphite App
arsenm
requested review from
andykaylor,
dtcxzyw,
jayfoad and
jcranmer-intel
llvmbot
added
the
llvm:analysis
Member
|
@llvm/pr-subscribers-llvm-analysis @llvm/pr-subscribers-llvm-ir Author: Matt Arsenault (arsenm) ChangesMultiply by large constant can be used to scale denormal inputs into The test cases are mostly the existing nofpclass test for ldexp, Full diff: https://github.com/llvm/llvm-project/pull/171730.diff 3 Files Affected:
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 9cb6f19b9340c..730b9a30c77ed 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -5646,6 +5646,24 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
KnownFPClass KnownLHS, KnownRHS;
computeKnownFPClass(Op->getOperand(1), DemandedElts, NeedForNan, KnownRHS,
Q, Depth + 1);
+
+ const APFloat *CRHS;
+ if (match(Op->getOperand(1), m_APFloat(CRHS))) {
+ // Match denormal scaling pattern, similar to the case in ldexp. If the
+ // constant's exponent is sufficiently large, the result cannot be
+ // subnormal.
+
+ // TODO: Should do general ConstantFPRange analysis.
+ const fltSemantics &Flt =
+ Op->getType()->getScalarType()->getFltSemantics();
+ unsigned Precision = APFloat::semanticsPrecision(Flt);
+ const int MantissaBits = Precision - 1;
+
+ int MinKnownExponent = ilogb(*CRHS);
+ if (MinKnownExponent >= MantissaBits)
+ Known.knownNot(fcSubnormal);
+ }
+
if (!KnownRHS.isKnownNeverNaN())
break;
diff --git a/llvm/test/Transforms/Attributor/nofpclass-fmul.ll b/llvm/test/Transforms/Attributor/nofpclass-fmul.ll
index 6d0edf0681ed4..0ad34105d4cb6 100644
--- a/llvm/test/Transforms/Attributor/nofpclass-fmul.ll
+++ b/llvm/test/Transforms/Attributor/nofpclass-fmul.ll
@@ -35,7 +35,7 @@ define float @ret_mul_exponent_f32_22(float %arg0) {
}
define float @ret_mul_exponent_f32_23(float %arg0) {
-; CHECK-LABEL: define float @ret_mul_exponent_f32_23(
+; CHECK-LABEL: define nofpclass(sub) float @ret_mul_exponent_f32_23(
; CHECK-SAME: float [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul float [[ARG0]], 0x4160000000000000
; CHECK-NEXT: ret float [[CALL]]
@@ -45,7 +45,7 @@ define float @ret_mul_exponent_f32_23(float %arg0) {
}
define float @ret_mul_exponent_f32_24(float %arg0) {
-; CHECK-LABEL: define float @ret_mul_exponent_f32_24(
+; CHECK-LABEL: define nofpclass(sub) float @ret_mul_exponent_f32_24(
; CHECK-SAME: float [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul float [[ARG0]], 0x4170000000000000
; CHECK-NEXT: ret float [[CALL]]
@@ -55,7 +55,7 @@ define float @ret_mul_exponent_f32_24(float %arg0) {
}
define float @ret_mul_exponent_f32_23_nnan(float nofpclass(nan) %arg0) {
-; CHECK-LABEL: define nofpclass(nan) float @ret_mul_exponent_f32_23_nnan(
+; CHECK-LABEL: define nofpclass(nan sub) float @ret_mul_exponent_f32_23_nnan(
; CHECK-SAME: float nofpclass(nan) [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul float [[ARG0]], 0x4160000000000000
; CHECK-NEXT: ret float [[CALL]]
@@ -85,7 +85,7 @@ define double @ret_mul_exponent_f64_51(double %arg0) {
}
define double @ret_mul_exponent_f64_52(double %arg0) {
-; CHECK-LABEL: define double @ret_mul_exponent_f64_52(
+; CHECK-LABEL: define nofpclass(sub) double @ret_mul_exponent_f64_52(
; CHECK-SAME: double [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul double [[ARG0]], 0x4330000000000000
; CHECK-NEXT: ret double [[CALL]]
@@ -95,7 +95,7 @@ define double @ret_mul_exponent_f64_52(double %arg0) {
}
define double @ret_mul_exponent_f64_53(double %arg0) {
-; CHECK-LABEL: define double @ret_mul_exponent_f64_53(
+; CHECK-LABEL: define nofpclass(sub) double @ret_mul_exponent_f64_53(
; CHECK-SAME: double [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul double [[ARG0]], 0x4340000000000000
; CHECK-NEXT: ret double [[CALL]]
@@ -125,7 +125,7 @@ define half @ret_mul_exponent_f16_9(half %arg0) {
}
define half @ret_mul_exponent_f16_10(half %arg0) {
-; CHECK-LABEL: define half @ret_mul_exponent_f16_10(
+; CHECK-LABEL: define nofpclass(sub) half @ret_mul_exponent_f16_10(
; CHECK-SAME: half [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul half [[ARG0]], 0xH6400
; CHECK-NEXT: ret half [[CALL]]
@@ -145,7 +145,7 @@ define bfloat @ret_mul_exponent_bf16_6(bfloat %arg0) {
}
define bfloat @ret_mul_exponent_bf16_7(bfloat %arg0) {
-; CHECK-LABEL: define bfloat @ret_mul_exponent_bf16_7(
+; CHECK-LABEL: define nofpclass(sub) bfloat @ret_mul_exponent_bf16_7(
; CHECK-SAME: bfloat [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul bfloat [[ARG0]], 0xR4300
; CHECK-NEXT: ret bfloat [[CALL]]
@@ -155,7 +155,7 @@ define bfloat @ret_mul_exponent_bf16_7(bfloat %arg0) {
}
define bfloat @ret_mul_exponent_bf16_8(bfloat %arg0) {
-; CHECK-LABEL: define bfloat @ret_mul_exponent_bf16_8(
+; CHECK-LABEL: define nofpclass(sub) bfloat @ret_mul_exponent_bf16_8(
; CHECK-SAME: bfloat [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul bfloat [[ARG0]], 0xR4380
; CHECK-NEXT: ret bfloat [[CALL]]
@@ -215,7 +215,7 @@ define float @ret_mul_exponent_f32_neg127(float %arg0) {
}
define <2 x float> @ret_mul_exponent_v2f32_splat_23(<2 x float> %arg0) {
-; CHECK-LABEL: define <2 x float> @ret_mul_exponent_v2f32_splat_23(
+; CHECK-LABEL: define nofpclass(sub) <2 x float> @ret_mul_exponent_v2f32_splat_23(
; CHECK-SAME: <2 x float> [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul <2 x float> [[ARG0]], splat (float 0x4160000000000000)
; CHECK-NEXT: ret <2 x float> [[CALL]]
@@ -267,7 +267,7 @@ define float @ret_mul_f32_0(float %arg0) {
}
define float @ret_mul_f32_inf(float %arg0) {
-; CHECK-LABEL: define float @ret_mul_f32_inf(
+; CHECK-LABEL: define nofpclass(sub) float @ret_mul_f32_inf(
; CHECK-SAME: float [[ARG0:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[CALL:%.*]] = fmul float [[ARG0]], 0x7FF0000000000000
; CHECK-NEXT: ret float [[CALL]]
diff --git a/llvm/test/Transforms/Attributor/nofpclass-nan-fmul.ll b/llvm/test/Transforms/Attributor/nofpclass-nan-fmul.ll
index fbf6c2e0981fb..f6bb1354cb74f 100644
--- a/llvm/test/Transforms/Attributor/nofpclass-nan-fmul.ll
+++ b/llvm/test/Transforms/Attributor/nofpclass-nan-fmul.ll
@@ -195,7 +195,7 @@ define float @ret_fmul_square_nnan_nzero(float nofpclass(nan zero) %arg) #0 {
}
define float @ret_fmul_ieee_inf(float %arg) {
-; CHECK-LABEL: define float @ret_fmul_ieee_inf
+; CHECK-LABEL: define nofpclass(sub) float @ret_fmul_ieee_inf
; CHECK-SAME: (float [[ARG:%.*]]) #[[ATTR4:[0-9]+]] {
; CHECK-NEXT: [[FMUL:%.*]] = fmul float [[ARG]], 0x7FF0000000000000
; CHECK-NEXT: ret float [[FMUL]]
|
arsenm
force-pushed
the
users/arsenm/valuetracking/fmul-large-exponent-constant-is-not-subnormal
branch
from
fed7047 to
b552c70
Compare
arsenm
force-pushed
the
users/arsenm/valuetracking/add-baseline-test-fmul-exponent-handling
branch
from
89db664 to
ef8a54e
Compare
Base automatically changed from
users/arsenm/valuetracking/add-baseline-test-fmul-exponent-handling
to
main
December 11, 2025 12:06
…rmals Multiply by large constant can be used to scale denormal inputs into a normal range. This pattern appears frequently in math function library implementations to make use of hardware instructions that do not support denormals. We already handle this case for ldexp, but now canonicalize ldexp by a constant to an fmul. The test cases are mostly the existing nofpclass test for ldexp, run through the new instcombine to replace ldexp with fmul.
arsenm
force-pushed
the
users/arsenm/valuetracking/fmul-large-exponent-constant-is-not-subnormal
branch
from
b552c70 to
892f156
Compare
arsenm
deleted the
users/arsenm/valuetracking/fmul-large-exponent-constant-is-not-subnormal
branch
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.
Multiply by large constant can be used to scale denormal inputs into
a normal range. This pattern appears frequently in math function library
implementations to make use of hardware instructions that do not support
denormals. We already handle this case for ldexp, but now canonicalize
ldexp by a constant to an fmul.
The test cases are mostly the existing nofpclass test for ldexp,
run through the new instcombine to replace ldexp with fmul.