Place fold at the end of visitSTORE

Author: abhishek-kaushik22

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -22541,9 +22541,6 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) {
   SDValue Value = ST->getValue();
   SDValue Ptr   = ST->getBasePtr();
 
-  if (SDValue MaskedStore = foldToMaskedStore(ST, DAG, SDLoc(N)))
-    return MaskedStore;
-
   // If this is a store of a bit convert, store the input value if the
   // resultant store does not need a higher alignment than the original.
   if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
@@ -22772,6 +22769,9 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) {
   if (SDValue NewSt = splitMergedValStore(ST))
     return NewSt;
 
+  if (SDValue MaskedStore = foldToMaskedStore(ST, DAG, SDLoc(N)))
+    return MaskedStore;
+
   return ReduceLoadOpStoreWidth(N);
 }
 

llvm/test/CodeGen/AArch64/combine-storetomstore.ll

@@ -1117,11 +1117,11 @@ define void @test_masked_store_unaligned_v8i32(<8 x i32> %data, ptr %ptr, <8 x i
 ; SVE-LABEL: test_masked_store_unaligned_v8i32:
 ; SVE:       // %bb.0:
 ; SVE-NEXT:    // kill: def $q0 killed $q0 def $z0
-; SVE-NEXT:    zip2 v3.8b, v2.8b, v0.8b
-; SVE-NEXT:    zip1 v2.8b, v2.8b, v0.8b
-; SVE-NEXT:    add x8, x0, #17
+; SVE-NEXT:    zip1 v3.8b, v2.8b, v0.8b
+; SVE-NEXT:    zip2 v2.8b, v2.8b, v0.8b
+; SVE-NEXT:    add x8, x0, #1
 ; SVE-NEXT:    ptrue p0.s, vl4
-; SVE-NEXT:    add x9, x0, #1
+; SVE-NEXT:    add x9, x0, #17
 ; SVE-NEXT:    // kill: def $q1 killed $q1 def $z1
 ; SVE-NEXT:    ushll v3.4s, v3.4h, #0
 ; SVE-NEXT:    ushll v2.4s, v2.4h, #0
@@ -1131,8 +1131,8 @@ define void @test_masked_store_unaligned_v8i32(<8 x i32> %data, ptr %ptr, <8 x i
 ; SVE-NEXT:    cmlt v2.4s, v2.4s, #0
 ; SVE-NEXT:    cmpne p1.s, p0/z, z3.s, #0
 ; SVE-NEXT:    cmpne p0.s, p0/z, z2.s, #0
-; SVE-NEXT:    st1w { z1.s }, p1, [x8]
-; SVE-NEXT:    st1w { z0.s }, p0, [x9]
+; SVE-NEXT:    st1w { z0.s }, p1, [x8]
+; SVE-NEXT:    st1w { z1.s }, p0, [x9]
 ; SVE-NEXT:    ret
   %ptr_i8 = getelementptr i8, ptr %ptr, i32 1
   %ptr_vec = bitcast ptr %ptr_i8 to ptr

llvm/test/CodeGen/X86/combine-storetomstore.ll

@@ -392,8 +392,8 @@ define void @test_masked_store_success_v8i64(<8 x i64> %x, ptr %ptr, <8 x i1> %m
 ; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
 ; AVX-NEXT:    vpmovsxdq %xmm2, %xmm2
 ; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm4, %ymm2
-; AVX-NEXT:    vmaskmovpd %ymm1, %ymm3, 32(%rdi)
 ; AVX-NEXT:    vmaskmovpd %ymm0, %ymm2, (%rdi)
+; AVX-NEXT:    vmaskmovpd %ymm1, %ymm3, 32(%rdi)
 ; AVX-NEXT:    vzeroupper
 ; AVX-NEXT:    retq
 ;
@@ -405,8 +405,8 @@ define void @test_masked_store_success_v8i64(<8 x i64> %x, ptr %ptr, <8 x i1> %m
 ; AVX2-NEXT:    vpmovzxwd {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
 ; AVX2-NEXT:    vpslld $31, %xmm2, %xmm2
 ; AVX2-NEXT:    vpmovsxdq %xmm2, %ymm2
-; AVX2-NEXT:    vpmaskmovq %ymm1, %ymm3, 32(%rdi)
 ; AVX2-NEXT:    vpmaskmovq %ymm0, %ymm2, (%rdi)
+; AVX2-NEXT:    vpmaskmovq %ymm1, %ymm3, 32(%rdi)
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 ;
@@ -502,8 +502,8 @@ define void @test_masked_store_success_v8f64(<8 x double> %x, ptr %ptr, <8 x i1>
 ; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
 ; AVX-NEXT:    vpmovsxdq %xmm2, %xmm2
 ; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm4, %ymm2
-; AVX-NEXT:    vmaskmovpd %ymm1, %ymm3, 32(%rdi)
 ; AVX-NEXT:    vmaskmovpd %ymm0, %ymm2, (%rdi)
+; AVX-NEXT:    vmaskmovpd %ymm1, %ymm3, 32(%rdi)
 ; AVX-NEXT:    vzeroupper
 ; AVX-NEXT:    retq
 ;
@@ -515,8 +515,8 @@ define void @test_masked_store_success_v8f64(<8 x double> %x, ptr %ptr, <8 x i1>
 ; AVX2-NEXT:    vpmovzxwd {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
 ; AVX2-NEXT:    vpslld $31, %xmm2, %xmm2
 ; AVX2-NEXT:    vpmovsxdq %xmm2, %ymm2
-; AVX2-NEXT:    vmaskmovpd %ymm1, %ymm3, 32(%rdi)
 ; AVX2-NEXT:    vmaskmovpd %ymm0, %ymm2, (%rdi)
+; AVX2-NEXT:    vmaskmovpd %ymm1, %ymm3, 32(%rdi)
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 ;
@@ -1506,8 +1506,8 @@ define void @test_masked_store_unaligned_v8i64(<8 x i64> %data, ptr %ptr, <8 x i
 ; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
 ; AVX-NEXT:    vpmovsxdq %xmm2, %xmm2
 ; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm4, %ymm2
-; AVX-NEXT:    vmaskmovpd %ymm1, %ymm3, 33(%rdi)
 ; AVX-NEXT:    vmaskmovpd %ymm0, %ymm2, 1(%rdi)
+; AVX-NEXT:    vmaskmovpd %ymm1, %ymm3, 33(%rdi)
 ; AVX-NEXT:    vzeroupper
 ; AVX-NEXT:    retq
 ;
@@ -1519,8 +1519,8 @@ define void @test_masked_store_unaligned_v8i64(<8 x i64> %data, ptr %ptr, <8 x i
 ; AVX2-NEXT:    vpmovzxwd {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
 ; AVX2-NEXT:    vpslld $31, %xmm2, %xmm2
 ; AVX2-NEXT:    vpmovsxdq %xmm2, %ymm2
-; AVX2-NEXT:    vpmaskmovq %ymm1, %ymm3, 33(%rdi)
 ; AVX2-NEXT:    vpmaskmovq %ymm0, %ymm2, 1(%rdi)
+; AVX2-NEXT:    vpmaskmovq %ymm1, %ymm3, 33(%rdi)
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 ;