[llvm] [DAG] expandVecReduce - widen sub-legal vectors to not prematurely scalarize later reduction levels (PR #194672)

[Phoebe Wang via llvm-commits]

================
@@ -94,28 +100,38 @@ define i32 @extract_last_active_v2i32(<2 x i32> %a, <2 x i1> %c) {
 define i32 @extract_last_active_v3i32(<3 x i32> %a, <3 x i1> %c) {
 ; CHECK-LABEL: extract_last_active_v3i32:
 ; CHECK:       # %bb.0:
+; CHECK-NEXT:    movd %esi, %xmm2
+; CHECK-NEXT:    movd %edi, %xmm1
+; CHECK-NEXT:    movdqa %xmm1, %xmm3
+; CHECK-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; CHECK-NEXT:    movd %edx, %xmm2
+; CHECK-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm2[0]
 ; CHECK-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
-; CHECK-NEXT:    movd %edx, %xmm0
-; CHECK-NEXT:    movd %esi, %xmm1
-; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,0],xmm0[0,1]
-; CHECK-NEXT:    pslld $31, %xmm1
-; CHECK-NEXT:    psrad $31, %xmm1
-; CHECK-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; CHECK-NEXT:    pslld $31, %xmm3
+; CHECK-NEXT:    psrad $31, %xmm3
+; CHECK-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
+; CHECK-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
+; CHECK-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648,2147483648,2147483648]
+; CHECK-NEXT:    movdqa %xmm3, %xmm4
+; CHECK-NEXT:    por %xmm0, %xmm4
+; CHECK-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[2,3,2,3]
+; CHECK-NEXT:    por %xmm5, %xmm0
+; CHECK-NEXT:    pcmpgtd %xmm0, %xmm4
+; CHECK-NEXT:    pand %xmm4, %xmm3
+; CHECK-NEXT:    pandn %xmm5, %xmm4
+; CHECK-NEXT:    por %xmm3, %xmm4
+; CHECK-NEXT:    movd %xmm4, %eax
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,1,1]
 ; CHECK-NEXT:    movd %xmm0, %ecx
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; CHECK-NEXT:    movd %xmm0, %eax
 ; CHECK-NEXT:    cmpl %ecx, %eax
 ; CHECK-NEXT:    cmoval %eax, %ecx
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
-; CHECK-NEXT:    movd %xmm0, %eax
-; CHECK-NEXT:    cmpl %eax, %ecx
-; CHECK-NEXT:    cmovbel %eax, %ecx
-; CHECK-NEXT:    orl %esi, %edi
-; CHECK-NEXT:    orl %edx, %edi
-; CHECK-NEXT:    andb $1, %dil
+; CHECK-NEXT:    andl $3, %ecx
+; CHECK-NEXT:    por %xmm2, %xmm1
+; CHECK-NEXT:    movd %xmm1, %edx
+; CHECK-NEXT:    orl %esi, %edx
+; CHECK-NEXT:    andb $1, %dl
 ; CHECK-NEXT:    xorl %eax, %eax
-; CHECK-NEXT:    cmpb $1, %dil
+; CHECK-NEXT:    cmpb $1, %dl
----------------
phoebewang wrote:

`isPow2VectorType` doesn't prevent here?

https://github.com/llvm/llvm-project/pull/194672