[llvm-branch-commits] [llvm] ValueTracking: Improve handling for fma/fmuladd (PR #175614)

Yingwei Zheng via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Tue Jan 13 09:10:04 PST 2026


================
@@ -84,11 +84,11 @@ define float @ret_fmuladd_different_same_arg_positive_addend(float noundef %arg0
 
 ; 1. operand0=positive, operand1=negative, operand2=positive
 define half @ret_fma__pos0__neg1__pos2() {
-; CHECK-LABEL: define half @ret_fma__pos0__neg1__pos2() {
+; CHECK-LABEL: define nofpclass(nzero) half @ret_fma__pos0__neg1__pos2() {
----------------
dtcxzyw wrote:

Alive2 also complaints about this transform with fmuladd. Adding `nofpclass(nzero)` only works for the fmul+fadd variant. If the inferred flags don't work for both variants, we must rewrite fmuladd into one specific variant.



https://github.com/llvm/llvm-project/pull/175614


More information about the llvm-branch-commits mailing list