[llvm] [AMDGPU] Fix sincos insertion point when argument is a phi (PR #215958)

[Arseniy Obolenskiy via llvm-commits]

https://github.com/aobolensk created https://github.com/llvm/llvm-project/pull/215958

Fix insertion landed between phis when the argument wasn't the last one

>From 351b6cd768cb24ff843688ca6d9cbc81ce000e80 Mon Sep 17 00:00:00 2001
From: Arseniy Obolenskiy <arseniy.obolenskiy at amd.com>
Date: Thu, 13 Aug 2026 08:00:38 +0200
Subject: [PATCH] [AMDGPU] Fix sincos insertion point when argument is a phi

Fix insertion landed between phis when the argument wasn't the last one
---
 llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp     |  7 +++-
 .../amdgpu-simplify-libcall-sincos.defined.ll | 35 +++++++++++++++++++
 2 files changed, 41 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp b/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp
index 7291fdaca0d69..996127526c222 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp
@@ -1681,7 +1681,7 @@ AMDGPULibCalls::insertSinCos(Value *Arg, FastMathFlags FMF, IRBuilder<> &B,
     // sincos call there. Otherwise, right after the allocas works well enough
     // if it's an argument or constant.
 
-    B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
+    B.SetInsertPoint(*ArgInst->getInsertionPointAfterDef());
 
     // SetInsertPoint unwelcomely always tries to set the debug loc.
     B.SetCurrentDebugLocation(DL);
@@ -1794,6 +1794,11 @@ bool AMDGPULibCalls::fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B,
   if (SinCalls.empty() || CosCalls.empty())
     return false;
 
+  // insertSinCos needs an insertion point after the argument's def.
+  if (auto *ArgInst = dyn_cast<Instruction>(CArgVal);
+      ArgInst && !ArgInst->getInsertionPointAfterDef())
+    return false;
+
   B.setFastMathFlags(FMF);
   B.setDefaultFPMathTag(FPMath);
   DILocation *DbgLoc = DILocation::getMergedLocations(MergeDbgLocs);
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-sincos.defined.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-sincos.defined.ll
index 037ae2897a22f..7a673dd56653b 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-sincos.defined.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-sincos.defined.ll
@@ -119,6 +119,41 @@ entry:
   ret void
 }
 
+; The merged call must go after the whole phi group, not right after %x.
+define float @sincos_f32_arg_is_not_last_phi(i1 %c, float %a, float %b, float %d) {
+; CHECK-LABEL: define float @sincos_f32_arg_is_not_last_phi
+; CHECK-SAME: (i1 [[C:%.*]], float [[A:%.*]], float [[B:%.*]], float [[D:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
+; CHECK-NEXT:    br i1 [[C]], label [[T:%.*]], label [[M:%.*]]
+; CHECK:       t:
+; CHECK-NEXT:    br label [[M]]
+; CHECK:       m:
+; CHECK-NEXT:    [[X:%.*]] = phi float [ [[A]], [[T]] ], [ [[B]], [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[Y:%.*]] = phi float [ [[D]], [[T]] ], [ 0.000000e+00, [[ENTRY]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
+; CHECK-NEXT:    [[COS:%.*]] = call contract float @_Z3cosf(float [[X]])
+; CHECK-NEXT:    [[SUM:%.*]] = fadd float [[TMP0]], [[TMP1]]
+; CHECK-NEXT:    [[SUM2:%.*]] = fadd float [[SUM]], [[Y]]
+; CHECK-NEXT:    ret float [[SUM2]]
+;
+entry:
+  br i1 %c, label %t, label %m
+
+t:
+  br label %m
+
+m:
+  %x = phi float [ %a, %t ], [ %b, %entry ]
+  %y = phi float [ %d, %t ], [ 0.0, %entry ]
+  %sin = call contract float @_Z3sinf(float %x)
+  %cos = call contract float @_Z3cosf(float %x)
+  %sum = fadd float %sin, %cos
+  %sum2 = fadd float %sum, %y
+  ret float %sum2
+}
+
 define void @sincos_f32_value_is_same_constantfp(ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
 ; CHECK-LABEL: define void @sincos_f32_value_is_same_constantfp
 ; CHECK-SAME: (ptr addrspace(1) writeonly captures(none) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) [[COS_OUT:%.*]]) {