[llvm] Update foldFMulReassoc to respect absent fast-math flags (PR #88589)

[Andy Kaylor via llvm-commits]

================
@@ -652,12 +652,49 @@ define float @fdiv_constant_numerator_fmul(float %x) {
 ; CHECK-LABEL: @fdiv_constant_numerator_fmul(
 ; CHECK-NEXT:    [[T3:%.*]] = fdiv reassoc float 1.200000e+07, [[X:%.*]]
 ; CHECK-NEXT:    ret float [[T3]]
+;
+  %t1 = fdiv reassoc float 2.0e+3, %x
+  %t3 = fmul reassoc float %t1, 6.0e+3
+  ret float %t3
+}
+
+; C1/X * C2 => (C1*C2) / X with mixed fast-math flags
+
+define float @fdiv_constant_numerator_fmul_mixed(float %x) {
+; CHECK-LABEL: @fdiv_constant_numerator_fmul_mixed(
+; CHECK-NEXT:    [[T3:%.*]] = fdiv reassoc float 1.200000e+07, [[X:%.*]]
+; CHECK-NEXT:    ret float [[T3]]
+;
+  %t1 = fdiv reassoc float 2.0e+3, %x
+  %t3 = fmul fast float %t1, 6.0e+3
+  ret float %t3
+}
+
+; C1/X * C2 => (C1*C2) / X with full fast-math flags
+
+define float @fdiv_constant_numerator_fmul_fast(float %x) {
+; CHECK-LABEL: @fdiv_constant_numerator_fmul_fast(
+; CHECK-NEXT:    [[T3:%.*]] = fdiv fast float 1.200000e+07, [[X:%.*]]
+; CHECK-NEXT:    ret float [[T3]]
+;
+  %t1 = fdiv fast float 2.0e+3, %x
+  %t3 = fmul fast float %t1, 6.0e+3
+  ret float %t3
+}
+
+; C1/X * C2 => (C1*C2) / X with no fast-math flags on the fdiv
+
+define float @fdiv_constant_numerator_fmul_precdiv(float %x) {
+; CHECK-LABEL: @fdiv_constant_numerator_fmul_precdiv(
+; CHECK-NEXT:    [[T4:%.*]] = fdiv reassoc float 1.200000e+07, [[X:%.*]]
+; CHECK-NEXT:    ret float [[T4]]
----------------
andykaylor wrote:

Yes, it was being folded by an additional case that handled non-constant fdiv and fmul. I've fixed that now. It required updating the input IR for a lot of other cases. In a couple cases adding the extra flags triggered a different transformation than we were previously testing for.

https://github.com/llvm/llvm-project/pull/88589