[X86] combineConcatVectorOps - add handling to concat RCPPS/RSQRTPS intrinsics together

[RKSimon]

Limited to 128->256 cases as we can't safely convert to the RCP14/RSQRT14 variants

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

Changes

Limited to 128->256 cases as we can't safely convert to the RCP14/RSQRT14 variants

---

Full diff: https://github.com/llvm/llvm-project/pull/170148.diff

3 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+6)
• (modified) llvm/test/CodeGen/X86/combine-rcp.ll (+12-12)
• (modified) llvm/test/CodeGen/X86/combine-rsqrt.ll (+12-12)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 9e20807ddf304..74a02711bd98a 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -59464,6 +59464,12 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
         return DAG.getNode(Opcode, DL, VT, ConcatSubOperand(VT, Ops, 0));
       }
       break;
+    case X86ISD::FRCP:
+    case X86ISD::FRSQRT:
+      if (!IsSplat && VT.is256BitVector()) {
+        return DAG.getNode(Opcode, DL, VT, ConcatSubOperand(VT, Ops, 0));
+      }
+      break;
     case X86ISD::HADD:
     case X86ISD::HSUB:
     case X86ISD::FHADD:
diff --git a/llvm/test/CodeGen/X86/combine-rcp.ll b/llvm/test/CodeGen/X86/combine-rcp.ll
index 7de3e96d592db..4647516528bf3 100644
--- a/llvm/test/CodeGen/X86/combine-rcp.ll
+++ b/llvm/test/CodeGen/X86/combine-rcp.ll
@@ -14,9 +14,9 @@ define <8 x float> @concat_rcp_v8f32_v4f32(<4 x float> %a0, <4 x float> %a1) {
 ;
 ; AVX-LABEL: concat_rcp_v8f32_v4f32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vrcpps %xmm0, %xmm0
-; AVX-NEXT:    vrcpps %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vrcpps %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <4 x float> @llvm.x86.sse.rcp.ps(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.x86.sse.rcp.ps(<4 x float> %a1)
@@ -36,23 +36,23 @@ define <16 x float> @concat_rcp_v16f32_v4f32(<4 x float> %a0, <4 x float> %a1, <
 ;
 ; AVX1OR2-LABEL: concat_rcp_v16f32_v4f32:
 ; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vrcpps %xmm0, %xmm0
-; AVX1OR2-NEXT:    vrcpps %xmm1, %xmm1
-; AVX1OR2-NEXT:    vrcpps %xmm2, %xmm2
-; AVX1OR2-NEXT:    vrcpps %xmm3, %xmm3
+; AVX1OR2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX1OR2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1OR2-NEXT:    vrcpps %ymm0, %ymm0
 ; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1OR2-NEXT:    vrcpps %ymm1, %ymm1
 ; AVX1OR2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_rcp_v16f32_v4f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vrcpps %xmm0, %xmm0
-; AVX512-NEXT:    vrcpps %xmm1, %xmm1
-; AVX512-NEXT:    vrcpps %xmm2, %xmm2
-; AVX512-NEXT:    vrcpps %xmm3, %xmm3
-; AVX512-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX512-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX512-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX512-NEXT:    vinsertf64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-NEXT:    vrcpps %ymm0, %ymm0
+; AVX512-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX512-NEXT:    vrcpps %ymm1, %ymm1
+; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x float> @llvm.x86.sse.rcp.ps(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.x86.sse.rcp.ps(<4 x float> %a1)
diff --git a/llvm/test/CodeGen/X86/combine-rsqrt.ll b/llvm/test/CodeGen/X86/combine-rsqrt.ll
index 78688701f8cd3..b373458654419 100644
--- a/llvm/test/CodeGen/X86/combine-rsqrt.ll
+++ b/llvm/test/CodeGen/X86/combine-rsqrt.ll
@@ -14,9 +14,9 @@ define <8 x float> @concat_rsqrt_v8f32_v4f32(<4 x float> %a0, <4 x float> %a1) {
 ;
 ; AVX-LABEL: concat_rsqrt_v8f32_v4f32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vrsqrtps %xmm0, %xmm0
-; AVX-NEXT:    vrsqrtps %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vrsqrtps %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float> %a1)
@@ -36,23 +36,23 @@ define <16 x float> @concat_rsqrt_v16f32_v4f32(<4 x float> %a0, <4 x float> %a1,
 ;
 ; AVX1OR2-LABEL: concat_rsqrt_v16f32_v4f32:
 ; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vrsqrtps %xmm0, %xmm0
-; AVX1OR2-NEXT:    vrsqrtps %xmm1, %xmm1
-; AVX1OR2-NEXT:    vrsqrtps %xmm2, %xmm2
-; AVX1OR2-NEXT:    vrsqrtps %xmm3, %xmm3
+; AVX1OR2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX1OR2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1OR2-NEXT:    vrsqrtps %ymm0, %ymm0
 ; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1OR2-NEXT:    vrsqrtps %ymm1, %ymm1
 ; AVX1OR2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_rsqrt_v16f32_v4f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vrsqrtps %xmm0, %xmm0
-; AVX512-NEXT:    vrsqrtps %xmm1, %xmm1
-; AVX512-NEXT:    vrsqrtps %xmm2, %xmm2
-; AVX512-NEXT:    vrsqrtps %xmm3, %xmm3
-; AVX512-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX512-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX512-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX512-NEXT:    vinsertf64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-NEXT:    vrsqrtps %ymm0, %ymm0
+; AVX512-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX512-NEXT:    vrsqrtps %ymm1, %ymm1
+; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float> %a1)