[llvm] [DAG] SimplifyDemandedVectorElts - add handling for INT<->FP conversions (PR #117884)

[Simon Pilgrim via llvm-commits]

================
@@ -141,56 +141,61 @@ declare <8 x half> @llvm.ldexp.v8f16.v8i16(<8 x half>, <8 x i16>)
 define <8 x half> @fmul_pow2_8xhalf(<8 x i16> %i) {
 ; CHECK-SSE-LABEL: fmul_pow2_8xhalf:
 ; CHECK-SSE:       # %bb.0:
-; CHECK-SSE-NEXT:    subq $88, %rsp
-; CHECK-SSE-NEXT:    .cfi_def_cfa_offset 96
+; CHECK-SSE-NEXT:    subq $104, %rsp
+; CHECK-SSE-NEXT:    .cfi_def_cfa_offset 112
 ; CHECK-SSE-NEXT:    movdqa %xmm0, %xmm1
 ; CHECK-SSE-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4,4,5,5,6,6,7,7]
 ; CHECK-SSE-NEXT:    pslld $23, %xmm1
 ; CHECK-SSE-NEXT:    movdqa {{.*#+}} xmm2 = [1065353216,1065353216,1065353216,1065353216]
 ; CHECK-SSE-NEXT:    paddd %xmm2, %xmm1
 ; CHECK-SSE-NEXT:    cvttps2dq %xmm1, %xmm1
-; CHECK-SSE-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill
+; CHECK-SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; CHECK-SSE-NEXT:    pslld $16, %xmm1
+; CHECK-SSE-NEXT:    movdqa %xmm1, (%rsp) # 16-byte Spill
 ; CHECK-SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
 ; CHECK-SSE-NEXT:    pslld $23, %xmm0
 ; CHECK-SSE-NEXT:    paddd %xmm2, %xmm0
 ; CHECK-SSE-NEXT:    cvttps2dq %xmm0, %xmm0
+; CHECK-SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
 ; CHECK-SSE-NEXT:    pslld $16, %xmm0
-; CHECK-SSE-NEXT:    psrld $16, %xmm0
 ; CHECK-SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; CHECK-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; CHECK-SSE-NEXT:    psrld $16, %xmm0
 ; CHECK-SSE-NEXT:    cvtdq2ps %xmm0, %xmm0
 ; CHECK-SSE-NEXT:    callq __truncsfhf2 at PLT
 ; CHECK-SSE-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; CHECK-SSE-NEXT:    cvtdq2ps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; CHECK-SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; CHECK-SSE-NEXT:    psrlq $48, %xmm0
+; CHECK-SSE-NEXT:    cvtdq2ps %xmm0, %xmm0
----------------
RKSimon wrote:

The bigger problem appears to be all these cvtdq2ps(shuffle) calls that could be replaced with shuffle(cvtdq2ps) and share the same cvtdq2ps

https://github.com/llvm/llvm-project/pull/117884