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

[Phoebe Wang via llvm-commits]

================
@@ -49,17 +50,22 @@ define i32 @extract_last_active_v4i32_no_default(<4 x i32> %a, <4 x i1> %c) {
 ; CHECK-NEXT:    psrad $31, %xmm1
 ; CHECK-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
 ; CHECK-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; CHECK-NEXT:    movd %xmm0, %eax
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; CHECK-NEXT:    movd %xmm0, %ecx
-; CHECK-NEXT:    cmpl %eax, %ecx
-; CHECK-NEXT:    cmoval %ecx, %eax
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
+; CHECK-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648,2147483648,2147483648]
+; CHECK-NEXT:    movdqa %xmm1, %xmm2
+; CHECK-NEXT:    por %xmm0, %xmm2
+; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]
+; CHECK-NEXT:    por %xmm3, %xmm0
+; CHECK-NEXT:    pcmpgtd %xmm0, %xmm2
+; CHECK-NEXT:    pand %xmm2, %xmm1
+; CHECK-NEXT:    pandn %xmm3, %xmm2
+; CHECK-NEXT:    por %xmm1, %xmm2
+; CHECK-NEXT:    movd %xmm2, %eax
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
 ; CHECK-NEXT:    movd %xmm0, %ecx
 ; CHECK-NEXT:    cmpl %ecx, %eax
-; CHECK-NEXT:    cmovbel %ecx, %eax
-; CHECK-NEXT:    movl -24(%rsp,%rax,4), %eax
+; CHECK-NEXT:    cmoval %eax, %ecx
+; CHECK-NEXT:    andl $3, %ecx
+; CHECK-NEXT:    movl -24(%rsp,%rcx,4), %eax
----------------
phoebewang wrote:

5 more instructions here.

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