[llvm] [X86] Prefer 128-bit vectors for Bulldozer CPUs (PR #225838)

[Simon Pilgrim via llvm-commits]

================
@@ -87,13 +87,13 @@ define void @fma_4f64() #0 {
 ; NO-FMA-NEXT:    store double [[FMA3]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 3), align 8
 ; NO-FMA-NEXT:    ret void
 ;
-; FMA-LABEL: @fma_4f64(
-; FMA-NEXT:    [[TMP1:%.*]] = load <4 x double>, ptr @srcA64, align 8
-; FMA-NEXT:    [[TMP2:%.*]] = load <4 x double>, ptr @srcB64, align 8
-; FMA-NEXT:    [[TMP3:%.*]] = load <4 x double>, ptr @srcC64, align 8
-; FMA-NEXT:    [[TMP4:%.*]] = call <4 x double> @llvm.fma.v4f64(<4 x double> [[TMP1]], <4 x double> [[TMP2]], <4 x double> [[TMP3]])
-; FMA-NEXT:    store <4 x double> [[TMP4]], ptr @dst64, align 8
-; FMA-NEXT:    ret void
+; FMA512-LABEL: @fma_4f64(
----------------
RKSimon wrote:

missing FMA256? you need  to add the missing FMA check prefix and regenerate

https://github.com/llvm/llvm-project/pull/225838