[HEXAGON] Fix semantics of ordered FP compares #131089
Merged
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
For the ordered FP compare bitcode instructions, the Hexagon backend was assuming that no operand could be a NaN. This assumption is flawed. This patch fixes the code-generation to produce fpcmp.uo and and appropriate bit comparison operators to account for the case when an operand to a FP compare is a NaN. Change-Id: I2c19c96d5bf326af73433f94c2f955ea426e8b6e
|
@llvm/pr-subscribers-backend-hexagon Author: None (aankit-ca) ChangesFor the ordered FP compare bitcode instructions, the Hexagon backend was assuming that no operand could be a NaN. This assumption is flawed. This patch fixes the code-generation to produce fpcmp.uo and and appropriate bit comparison operators to account for the case when an operand to a FP compare is a NaN. Full diff: https://github.com/llvm/llvm-project/pull/131089.diff 2 Files Affected:
diff --git a/llvm/lib/Target/Hexagon/HexagonPatterns.td b/llvm/lib/Target/Hexagon/HexagonPatterns.td
index cba5ff1ab0d9b..244f204539c89 100644
--- a/llvm/lib/Target/Hexagon/HexagonPatterns.td
+++ b/llvm/lib/Target/Hexagon/HexagonPatterns.td
@@ -721,11 +721,6 @@ def: OpR_RR_pat<A2_vcmpwgtu, setugt, v2i1, V2I32>;
def: OpR_RR_pat<F2_sfcmpeq, seteq, i1, F32>;
def: OpR_RR_pat<F2_sfcmpgt, setgt, i1, F32>;
def: OpR_RR_pat<F2_sfcmpge, setge, i1, F32>;
-def: OpR_RR_pat<F2_sfcmpeq, setoeq, i1, F32>;
-def: OpR_RR_pat<F2_sfcmpgt, setogt, i1, F32>;
-def: OpR_RR_pat<F2_sfcmpge, setoge, i1, F32>;
-def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setolt>, i1, F32>;
-def: OpR_RR_pat<F2_sfcmpge, RevCmp<setole>, i1, F32>;
def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setlt>, i1, F32>;
def: OpR_RR_pat<F2_sfcmpge, RevCmp<setle>, i1, F32>;
def: OpR_RR_pat<F2_sfcmpuo, setuo, i1, F32>;
@@ -733,11 +728,6 @@ def: OpR_RR_pat<F2_sfcmpuo, setuo, i1, F32>;
def: OpR_RR_pat<F2_dfcmpeq, seteq, i1, F64>;
def: OpR_RR_pat<F2_dfcmpgt, setgt, i1, F64>;
def: OpR_RR_pat<F2_dfcmpge, setge, i1, F64>;
-def: OpR_RR_pat<F2_dfcmpeq, setoeq, i1, F64>;
-def: OpR_RR_pat<F2_dfcmpgt, setogt, i1, F64>;
-def: OpR_RR_pat<F2_dfcmpge, setoge, i1, F64>;
-def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setolt>, i1, F64>;
-def: OpR_RR_pat<F2_dfcmpge, RevCmp<setole>, i1, F64>;
def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setlt>, i1, F64>;
def: OpR_RR_pat<F2_dfcmpge, RevCmp<setle>, i1, F64>;
def: OpR_RR_pat<F2_dfcmpuo, setuo, i1, F64>;
@@ -900,15 +890,35 @@ def: OpmR_RR_pat<Cmpud<F2_dfcmpge>, RevCmp<setule>, i1, F64>;
def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>, RevCmp<setult>, i1, F64>;
def: OpmR_RR_pat<Cmpudn<F2_dfcmpeq>, setune, i1, F64>;
-def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setone, i1, F32>;
-def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setne, i1, F32>;
+class T4<InstHexagon MI1, InstHexagon MI2, InstHexagon MI3, InstHexagon MI4>
+ : OutPatFrag<(ops node:$Rs, node:$Rt),
+ (MI1 (MI2 (MI3 $Rs, $Rt), (MI4 $Rs, $Rt)))>;
-def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setone, i1, F64>;
-def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setne, i1, F64>;
+class Cmpof<InstHexagon MI>: T3<C2_andn, MI, F2_sfcmpuo>;
+class Cmpod<InstHexagon MI>: T3<C2_andn, MI, F2_dfcmpuo>;
+
+class Cmpofn<InstHexagon MI>: T4<C2_not, C2_or, MI, F2_sfcmpuo>;
+class Cmpodn<InstHexagon MI>: T4<C2_not, C2_or, MI, F2_dfcmpuo>;
+
+def: OpmR_RR_pat<Cmpof<F2_sfcmpeq>, setoeq, i1, F32>;
+def: OpmR_RR_pat<Cmpof<F2_sfcmpge>, setoge, i1, F32>;
+def: OpmR_RR_pat<Cmpof<F2_sfcmpgt>, setogt, i1, F32>;
+def: OpmR_RR_pat<Cmpof<F2_sfcmpge>, RevCmp<setole>, i1, F32>;
+def: OpmR_RR_pat<Cmpof<F2_sfcmpgt>, RevCmp<setolt>, i1, F32>;
+def: OpmR_RR_pat<Cmpofn<F2_sfcmpeq>, setone, i1, F32>;
+
+def: OpmR_RR_pat<Cmpod<F2_dfcmpeq>, setoeq, i1, F64>;
+def: OpmR_RR_pat<Cmpod<F2_dfcmpge>, setoge, i1, F64>;
+def: OpmR_RR_pat<Cmpod<F2_dfcmpgt>, setogt, i1, F64>;
+def: OpmR_RR_pat<Cmpod<F2_dfcmpge>, RevCmp<setole>, i1, F64>;
+def: OpmR_RR_pat<Cmpod<F2_dfcmpgt>, RevCmp<setolt>, i1, F64>;
+def: OpmR_RR_pat<Cmpodn<F2_dfcmpeq>, setone, i1, F64>;
def: OpmR_RR_pat<Outn<F2_sfcmpuo>, seto, i1, F32>;
def: OpmR_RR_pat<Outn<F2_dfcmpuo>, seto, i1, F64>;
+def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setne, i1, F32>;
+def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setne, i1, F64>;
// --(6) Select ----------------------------------------------------------
//
diff --git a/llvm/test/CodeGen/Hexagon/fcmp-nan.ll b/llvm/test/CodeGen/Hexagon/fcmp-nan.ll
new file mode 100644
index 0000000000000..1469402911601
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/fcmp-nan.ll
@@ -0,0 +1,189 @@
+; RUN: llc -march=hexagon < %s | FileCheck %s
+;
+; Test that all FP ordered compare instructions generate the correct
+; post-processing to accommodate NaNs.
+;
+; Specifically for ordered FP compares, we have to check if one of
+; the operands was a NaN to comform to the semantics of the ordered
+; fcmp bitcode instruction
+;
+target triple = "hexagon"
+
+;
+; Functions for float:
+;
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = sfcmp.eq(r0,r1)
+; CHECK-DAG: [[REG1:p([0-3])]] = sfcmp.uo(r0,r1)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_oeq_f(float %val, float %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp oeq float %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = sfcmp.eq(r0,r1)
+; CHECK-DAG: [[REG1:p([0-3])]] = sfcmp.uo(r0,r1)
+; CHECK: [[REG2:p([0-3])]] = or([[REG0]],[[REG1]])
+; CHECK: r0 = mux([[REG2]],#0,#1)
+;
+define i32 @compare_one_f(float %val, float %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp one float %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = sfcmp.gt(r0,r1)
+; CHECK-DAG: [[REG1:p([0-3])]] = sfcmp.uo(r0,r1)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_ogt_f(float %val, float %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp ogt float %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = sfcmp.ge(r1,r0)
+; CHECK-DAG: [[REG1:p([0-3])]] = sfcmp.uo(r1,r0)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_ole_f(float %val, float %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp ole float %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = sfcmp.ge(r0,r1)
+; CHECK-DAG: [[REG1:p([0-3])]] = sfcmp.uo(r0,r1)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_oge_f(float %val, float %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp oge float %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = sfcmp.gt(r1,r0)
+; CHECK-DAG: [[REG1:p([0-3])]] = sfcmp.uo(r1,r0)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_olt_f(float %val, float %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp olt float %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+
+;
+; Functions for double:
+;
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = dfcmp.eq(r1:0,r3:2)
+; CHECK-DAG: [[REG1:p([0-3])]] = dfcmp.uo(r1:0,r3:2)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_oeq_d(double %val, double %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp oeq double %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = dfcmp.eq(r1:0,r3:2)
+; CHECK-DAG: [[REG1:p([0-3])]] = dfcmp.uo(r1:0,r3:2)
+; CHECK: [[REG2:p([0-3])]] = or([[REG0]],[[REG1]])
+; CHECK: r0 = mux([[REG2]],#0,#1)
+;
+define i32 @compare_one_d(double %val, double %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp one double %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = dfcmp.gt(r1:0,r3:2)
+; CHECK-DAG: [[REG1:p([0-3])]] = dfcmp.uo(r1:0,r3:2)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_ogt_d(double %val, double %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp ogt double %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = dfcmp.ge(r3:2,r1:0)
+; CHECK-DAG: [[REG1:p([0-3])]] = dfcmp.uo(r3:2,r1:0)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_ole_d(double %val, double %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp ole double %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = dfcmp.ge(r1:0,r3:2)
+; CHECK-DAG: [[REG1:p([0-3])]] = dfcmp.uo(r1:0,r3:2)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_oge_d(double %val, double %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp oge double %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
+
+;
+; CHECK-DAG: [[REG0:p([0-3])]] = dfcmp.gt(r3:2,r1:0)
+; CHECK-DAG: [[REG1:p([0-3])]] = dfcmp.uo(r3:2,r1:0)
+; CHECK: [[REG2:p([0-3])]] = and([[REG0]],![[REG1]])
+; CHECK: r0 = mux([[REG2]],#1,#0)
+;
+define i32 @compare_olt_d(double %val, double %val2) local_unnamed_addr #0 {
+entry:
+ %cmpinf = fcmp olt double %val, %val2
+ %0 = zext i1 %cmpinf to i32
+ ret i32 %0
+}
+
|
androm3da
reviewed
Mar 13, 2025
androm3da
reviewed
Mar 13, 2025
iajbar
approved these changes
Mar 13, 2025
frederik-h
pushed a commit
to frederik-h/llvm-project
that referenced
this pull request
Mar 18, 2025
For the ordered FP compare bitcode instructions, the Hexagon backend was assuming that no operand could be a NaN. This assumption is flawed. This patch fixes the code-generation to produce fpcmp.uo and and appropriate bit comparison operators to account for the case when an operand to a FP compare is a NaN. Fix for llvm#129391 Co-authored-by: aankit-quic <[email protected]>
swift-ci
pushed a commit
to swiftlang/llvm-project
that referenced
this pull request
Mar 27, 2025
For the ordered FP compare bitcode instructions, the Hexagon backend was assuming that no operand could be a NaN. This assumption is flawed. This patch fixes the code-generation to produce fpcmp.uo and and appropriate bit comparison operators to account for the case when an operand to a FP compare is a NaN. Fix for llvm#129391 Co-authored-by: aankit-quic <[email protected]> (cherry picked from commit d642eec)
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.
For the ordered FP compare bitcode instructions, the Hexagon backend was assuming that no operand could be a NaN. This assumption is flawed. This patch fixes the code-generation to produce fpcmp.uo and and appropriate bit comparison operators to account for the case when an operand to a FP compare is a NaN.
Fix for #129391