[llvm] [InstCombine] Support folding intrinsics into phis (PR #151115)
Nikita Popov via llvm-commits
llvm-commits at lists.llvm.org
Tue Jul 29 02:55:03 PDT 2025
https://github.com/nikic created https://github.com/llvm/llvm-project/pull/151115
None
>From f1112e1e191ac266617521731f05164e8b15bbcc Mon Sep 17 00:00:00 2001
From: Nikita Popov <npopov at redhat.com>
Date: Tue, 29 Jul 2025 11:50:16 +0200
Subject: [PATCH] [InstCombine] Support folding intrinsics into phis
---
.../InstCombine/InstCombineCalls.cpp | 6 ++++-
.../InstCombine/fpclass-from-dom-cond.ll | 8 +++----
.../InstCombine/known-phi-recurse.ll | 5 +---
.../recurrence-binary-intrinsic.ll | 24 +++++++------------
4 files changed, 17 insertions(+), 26 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 1b78acea62bd6..e2007a40fe1f3 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -3897,10 +3897,14 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
// perform cross-lane operations.
if (isSafeToSpeculativelyExecuteWithVariableReplaced(&CI) &&
isNotCrossLaneOperation(II))
- for (Value *Op : II->args())
+ for (Value *Op : II->args()) {
if (auto *Sel = dyn_cast<SelectInst>(Op))
if (Instruction *R = FoldOpIntoSelect(*II, Sel))
return R;
+ if (auto *Phi = dyn_cast<PHINode>(Op))
+ if (Instruction *R = foldOpIntoPhi(*II, Phi))
+ return R;
+ }
if (Instruction *Shuf = foldShuffledIntrinsicOperands(II))
return Shuf;
diff --git a/llvm/test/Transforms/InstCombine/fpclass-from-dom-cond.ll b/llvm/test/Transforms/InstCombine/fpclass-from-dom-cond.ll
index 934852d1ca8bc..02042b102d61e 100644
--- a/llvm/test/Transforms/InstCombine/fpclass-from-dom-cond.ll
+++ b/llvm/test/Transforms/InstCombine/fpclass-from-dom-cond.ll
@@ -131,10 +131,10 @@ define i1 @test5(double %x, i1 %cond) {
; CHECK: if.then:
; CHECK-NEXT: ret i1 false
; CHECK: if.end:
+; CHECK-NEXT: [[TMP0:%.*]] = tail call i1 @llvm.is.fpclass.f64(double [[X]], i32 408)
; CHECK-NEXT: br label [[EXIT]]
; CHECK: exit:
-; CHECK-NEXT: [[Y:%.*]] = phi double [ -1.000000e+00, [[ENTRY:%.*]] ], [ [[X]], [[IF_END]] ]
-; CHECK-NEXT: [[RET:%.*]] = tail call i1 @llvm.is.fpclass.f64(double [[Y]], i32 408)
+; CHECK-NEXT: [[RET:%.*]] = phi i1 [ true, [[ENTRY:%.*]] ], [ [[TMP0]], [[IF_END]] ]
; CHECK-NEXT: ret i1 [[RET]]
;
entry:
@@ -391,11 +391,9 @@ define float @test_signbit_check_fail(float %x, i1 %cond) {
; CHECK: if.else:
; CHECK-NEXT: br i1 [[COND]], label [[IF_THEN2:%.*]], label [[IF_END]]
; CHECK: if.then2:
-; CHECK-NEXT: [[FNEG2:%.*]] = fneg float [[X]]
; CHECK-NEXT: br label [[IF_END]]
; CHECK: if.end:
-; CHECK-NEXT: [[VALUE:%.*]] = phi float [ [[FNEG]], [[IF_THEN1]] ], [ [[FNEG2]], [[IF_THEN2]] ], [ [[X]], [[IF_ELSE]] ]
-; CHECK-NEXT: [[RET:%.*]] = call float @llvm.fabs.f32(float [[VALUE]])
+; CHECK-NEXT: [[RET:%.*]] = phi float [ [[FNEG]], [[IF_THEN1]] ], [ [[X]], [[IF_THEN2]] ], [ [[X]], [[IF_ELSE]] ]
; CHECK-NEXT: ret float [[RET]]
;
%i32 = bitcast float %x to i32
diff --git a/llvm/test/Transforms/InstCombine/known-phi-recurse.ll b/llvm/test/Transforms/InstCombine/known-phi-recurse.ll
index c05cca93b035c..ac44e6c9df412 100644
--- a/llvm/test/Transforms/InstCombine/known-phi-recurse.ll
+++ b/llvm/test/Transforms/InstCombine/known-phi-recurse.ll
@@ -261,14 +261,11 @@ define i8 @knownbits_umax_select_test() {
; CHECK-NEXT: entry:
; CHECK-NEXT: br label [[LOOP:%.*]]
; CHECK: loop:
-; CHECK-NEXT: [[INDVAR:%.*]] = phi i8 [ 0, [[ENTRY:%.*]] ], [ [[CONTAIN:%.*]], [[LOOP]] ]
; CHECK-NEXT: [[COND0:%.*]] = call i1 @cond()
-; CHECK-NEXT: [[CONTAIN]] = call i8 @llvm.umax.i8(i8 [[INDVAR]], i8 1)
; CHECK-NEXT: [[COND1:%.*]] = call i1 @cond()
; CHECK-NEXT: br i1 [[COND1]], label [[EXIT:%.*]], label [[LOOP]]
; CHECK: exit:
-; CHECK-NEXT: [[BOOL:%.*]] = and i8 [[CONTAIN]], 1
-; CHECK-NEXT: ret i8 [[BOOL]]
+; CHECK-NEXT: ret i8 1
;
entry:
br label %loop
diff --git a/llvm/test/Transforms/InstCombine/recurrence-binary-intrinsic.ll b/llvm/test/Transforms/InstCombine/recurrence-binary-intrinsic.ll
index c637481ac9c6d..86e586ef0a16c 100644
--- a/llvm/test/Transforms/InstCombine/recurrence-binary-intrinsic.ll
+++ b/llvm/test/Transforms/InstCombine/recurrence-binary-intrinsic.ll
@@ -8,12 +8,11 @@ define i8 @simple_recurrence_intrinsic_smax(i8 %n, i8 %a, i8 %b) {
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i8 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT: [[SMAX_ACC:%.*]] = phi i8 [ [[SMAX:%.*]], %[[LOOP]] ], [ [[A]], %[[ENTRY]] ]
-; CHECK-NEXT: [[SMAX]] = call i8 @llvm.smax.i8(i8 [[SMAX_ACC]], i8 [[B]])
; CHECK-NEXT: [[IV_NEXT]] = add nuw i8 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i8 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]]
; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[SMAX:%.*]] = call i8 @llvm.smax.i8(i8 [[A]], i8 [[B]])
; CHECK-NEXT: ret i8 [[SMAX]]
;
entry:
@@ -38,12 +37,11 @@ define i8 @simple_recurrence_intrinsic_smin(i8 %n, i8 %a, i8 %b) {
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i8 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT: [[SMIN_ACC:%.*]] = phi i8 [ [[SMIN:%.*]], %[[LOOP]] ], [ [[A]], %[[ENTRY]] ]
-; CHECK-NEXT: [[SMIN]] = call i8 @llvm.smin.i8(i8 [[SMIN_ACC]], i8 [[B]])
; CHECK-NEXT: [[IV_NEXT]] = add nuw i8 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i8 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]]
; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[SMIN:%.*]] = call i8 @llvm.smin.i8(i8 [[A]], i8 [[B]])
; CHECK-NEXT: ret i8 [[SMIN]]
;
entry:
@@ -68,12 +66,11 @@ define i8 @simple_recurrence_intrinsic_umax(i8 %n, i8 %a, i8 %b) {
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i8 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT: [[UMAX_ACC:%.*]] = phi i8 [ [[UMAX:%.*]], %[[LOOP]] ], [ [[A]], %[[ENTRY]] ]
-; CHECK-NEXT: [[UMAX]] = call i8 @llvm.umax.i8(i8 [[UMAX_ACC]], i8 [[B]])
; CHECK-NEXT: [[IV_NEXT]] = add nuw i8 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i8 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]]
; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[UMAX:%.*]] = call i8 @llvm.umax.i8(i8 [[A]], i8 [[B]])
; CHECK-NEXT: ret i8 [[UMAX]]
;
entry:
@@ -98,12 +95,11 @@ define i8 @simple_recurrence_intrinsic_umin(i8 %n, i8 %a, i8 %b) {
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i8 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT: [[UMIN_ACC:%.*]] = phi i8 [ [[UMIN:%.*]], %[[LOOP]] ], [ [[A]], %[[ENTRY]] ]
-; CHECK-NEXT: [[UMIN]] = call i8 @llvm.umin.i8(i8 [[UMIN_ACC]], i8 [[B]])
; CHECK-NEXT: [[IV_NEXT]] = add nuw i8 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i8 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]]
; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[UMIN:%.*]] = call i8 @llvm.umin.i8(i8 [[A]], i8 [[B]])
; CHECK-NEXT: ret i8 [[UMIN]]
;
entry:
@@ -128,12 +124,11 @@ define float @simple_recurrence_intrinsic_maxnum(i32 %n, float %a, float %b) {
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT: [[FMAX_ACC:%.*]] = phi float [ [[FMAX:%.*]], %[[LOOP]] ], [ [[A]], %[[ENTRY]] ]
-; CHECK-NEXT: [[FMAX]] = call float @llvm.maxnum.f32(float [[FMAX_ACC]], float [[B]])
; CHECK-NEXT: [[IV_NEXT]] = add nuw i32 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]]
; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[FMAX:%.*]] = call float @llvm.maxnum.f32(float [[A]], float [[B]])
; CHECK-NEXT: ret float [[FMAX]]
;
entry:
@@ -157,12 +152,11 @@ define float @simple_recurrence_intrinsic_minnum(i32 %n, float %a, float %b) {
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT: [[FMIN_ACC:%.*]] = phi float [ [[FMIN:%.*]], %[[LOOP]] ], [ [[A]], %[[ENTRY]] ]
-; CHECK-NEXT: [[FMIN]] = call float @llvm.minnum.f32(float [[FMIN_ACC]], float [[B]])
; CHECK-NEXT: [[IV_NEXT]] = add nuw i32 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]]
; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[FMIN:%.*]] = call float @llvm.minnum.f32(float [[A]], float [[B]])
; CHECK-NEXT: ret float [[FMIN]]
;
entry:
@@ -186,12 +180,11 @@ define float @simple_recurrence_intrinsic_maximum(i32 %n, float %a, float %b) {
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT: [[FMAX_ACC:%.*]] = phi float [ [[FMAX:%.*]], %[[LOOP]] ], [ [[A]], %[[ENTRY]] ]
-; CHECK-NEXT: [[FMAX]] = call nnan float @llvm.maximum.f32(float [[FMAX_ACC]], float [[B]])
; CHECK-NEXT: [[IV_NEXT]] = add nuw i32 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]]
; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[FMAX:%.*]] = call nnan float @llvm.maximum.f32(float [[A]], float [[B]])
; CHECK-NEXT: ret float [[FMAX]]
;
entry:
@@ -215,12 +208,11 @@ define float @simple_recurrence_intrinsic_minimum(i32 %n, float %a, float %b) {
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT: [[FMIN_ACC:%.*]] = phi float [ [[FMIN:%.*]], %[[LOOP]] ], [ [[A]], %[[ENTRY]] ]
-; CHECK-NEXT: [[FMIN]] = call nnan float @llvm.minimum.f32(float [[FMIN_ACC]], float [[B]])
; CHECK-NEXT: [[IV_NEXT]] = add nuw i32 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]]
; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[FMIN:%.*]] = call nnan float @llvm.minimum.f32(float [[A]], float [[B]])
; CHECK-NEXT: ret float [[FMIN]]
;
entry:
More information about the llvm-commits
mailing list