[llvm] [SLP]Vectorize gathered loads (PR #107461)

[Simon Pilgrim via llvm-commits]

================
@@ -28,6 +38,19 @@ entry:
 }
 
 define void @vec3_fmuladd_64(ptr %Colour, double %0) {
+; NON-POWER-OF-2-LABEL: @vec3_fmuladd_64(
+; NON-POWER-OF-2-NEXT:  entry:
+; NON-POWER-OF-2-NEXT:    [[ARRAYIDX80:%.*]] = getelementptr float, ptr [[COLOUR:%.*]], i64 2
+; NON-POWER-OF-2-NEXT:    [[TMP1:%.*]] = insertelement <2 x double> poison, double [[TMP0:%.*]], i32 0
+; NON-POWER-OF-2-NEXT:    [[TMP2:%.*]] = shufflevector <2 x double> [[TMP1]], <2 x double> poison, <2 x i32> zeroinitializer
+; NON-POWER-OF-2-NEXT:    [[TMP3:%.*]] = call <2 x double> @llvm.fmuladd.v2f64(<2 x double> [[TMP2]], <2 x double> zeroinitializer, <2 x double> zeroinitializer)
+; NON-POWER-OF-2-NEXT:    [[TMP4:%.*]] = fptrunc <2 x double> [[TMP3]] to <2 x float>
+; NON-POWER-OF-2-NEXT:    store <2 x float> [[TMP4]], ptr [[COLOUR]], align 4
+; NON-POWER-OF-2-NEXT:    [[TMP5:%.*]] = call double @llvm.fmuladd.f64(double [[TMP0]], double 0.000000e+00, double 0.000000e+00)
+; NON-POWER-OF-2-NEXT:    [[CONV82:%.*]] = fptrunc double [[TMP5]] to float
+; NON-POWER-OF-2-NEXT:    store float [[CONV82]], ptr [[ARRAYIDX80]], align 4
+; NON-POWER-OF-2-NEXT:    ret void
+;
----------------
RKSimon wrote:

This appears to be the same as the CHECK case - maybe use CHECK as common prefix and add a POWER-OF-2 prefix?

https://github.com/llvm/llvm-project/pull/107461