[AArch64] Use custom lowering for {U,S}INT_TO_FP with i8.

fhahn · fhahn · commit 93bcb8342e73 · 2021-11-11T19:01:10.000+01:00
With fullfp16, it is cheaper to cast the {U,S}INT_TO_FP operand to i16 first, rather than promoting it to i32. The custom lowering for {U,S}INT_TO_FP already supports that, it just needs to be used. Reviewed By: dmgreen Differential Revision: https://reviews.llvm.org/D113601 (cherry-picked from c2ed9fd)
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -1003,10 +1003,6 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
     // elements smaller than i32, so promote the input to i32 first.
     setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v4i8, MVT::v4i32);
     setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v4i8, MVT::v4i32);
-    setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v8i8, MVT::v8i32);
-    setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v8i8, MVT::v8i32);
-    setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v16i8, MVT::v16i32);
-    setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v16i8, MVT::v16i32);
 
     // Similarly, there is no direct i32 -> f64 vector conversion instruction.
     setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
@@ -1019,13 +1015,21 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
 
     if (Subtarget->hasFullFP16()) {
+      setOperationAction(ISD::SINT_TO_FP, MVT::v8i8, Custom);
+      setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
+      setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Custom);
+      setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
       setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
       setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
       setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Custom);
       setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
     } else {
       // when AArch64 doesn't have fullfp16 support, promote the input
       // to i32 first.
+      setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v8i8, MVT::v8i32);
+      setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v8i8, MVT::v8i32);
+      setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v16i8, MVT::v16i32);
+      setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v16i8, MVT::v16i32);
       setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v4i16, MVT::v4i32);
       setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v4i16, MVT::v4i32);
       setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v8i16, MVT::v8i32);
diff --git a/llvm/test/CodeGen/AArch64/fp16-v8-instructions.ll b/llvm/test/CodeGen/AArch64/fp16-v8-instructions.ll
@@ -297,41 +297,51 @@ define <4 x half> @sitofp_v4i8(<4 x i8> %a) #0 {
 }
 
 define <8 x half> @sitofp_v8i8(<8 x i8> %a) #0 {
-; CHECK-LABEL: sitofp_v8i8:
-; CHECK-NEXT:    sshll v0.8h, v0.8b, #0
-; CHECK-NEXT:    sshll2 v1.4s, v0.8h, #0
-; CHECK-NEXT:    sshll v0.4s, v0.4h, #0
-; CHECK-NEXT:    scvtf v1.4s, v1.4s
-; CHECK-NEXT:    scvtf v0.4s, v0.4s
-; CHECK-NEXT:    fcvtn v1.4h, v1.4s
-; CHECK-NEXT:    fcvtn v0.4h, v0.4s
-; CHECK-NEXT:    mov v0.d[1], v1.d[0]
-; CHECK-NEXT:    ret
+; CHECK-CVT-LABEL: sitofp_v8i8:
+; CHECK-CVT-NEXT:    sshll v0.8h, v0.8b, #0
+; CHECK-CVT-NEXT:    sshll2 v1.4s, v0.8h, #0
+; CHECK-CVT-NEXT:    sshll v0.4s, v0.4h, #0
+; CHECK-CVT-NEXT:    scvtf v1.4s, v1.4s
+; CHECK-CVT-NEXT:    scvtf v0.4s, v0.4s
+; CHECK-CVT-NEXT:    fcvtn v1.4h, v1.4s
+; CHECK-CVT-NEXT:    fcvtn v0.4h, v0.4s
+; CHECK-CVT-NEXT:    mov v0.d[1], v1.d[0]
+; CHECK-CVT-NEXT:    ret
 ;
+; CHECK-FP16-LABEL: sitofp_v8i8:
+; CHECK-FP16-NEXT:    sshll v0.8h, v0.8b, #0
+; CHECK-FP16-NEXT:    scvtf v0.8h, v0.8h
+; CHECK-FP16-NEXT:    ret
   %1 = sitofp <8 x i8> %a to <8 x half>
   ret <8 x half> %1
 }
 
 define <16 x half> @sitofp_v16i8(<16 x i8> %a) #0 {
-; CHECK-LABEL: sitofp_v16i8:
-; CHECK-NEXT:    sshll2  v1.8h, v0.16b, #0
-; CHECK-NEXT:    sshll2  v2.4s, v1.8h, #0
-; CHECK-NEXT:    sshll   v1.4s, v1.4h, #0
-; CHECK-NEXT:    scvtf   v2.4s, v2.4s
-; CHECK-NEXT:    scvtf   v1.4s, v1.4s
-; CHECK-NEXT:    sshll   v0.8h, v0.8b, #0
-; CHECK-NEXT:    fcvtn   v2.4h, v2.4s
-; CHECK-NEXT:    fcvtn   v1.4h, v1.4s
-; CHECK-NEXT:    mov v1.d[1], v2.d[0]
-; CHECK-NEXT:    sshll2  v2.4s, v0.8h, #0
-; CHECK-NEXT:    sshll   v0.4s, v0.4h, #0
-; CHECK-NEXT:    scvtf   v2.4s, v2.4s
-; CHECK-NEXT:    scvtf   v0.4s, v0.4s
-; CHECK-NEXT:    fcvtn   v2.4h, v2.4s
-; CHECK-NEXT:    fcvtn   v0.4h, v0.4s
-; CHECK-NEXT:    mov v0.d[1], v2.d[0]
-; CHECK-NEXT:    ret
+; CHECK-CVT-LABEL: sitofp_v16i8:
+; CHECK-CVT-NEXT:    sshll2  v1.8h, v0.16b, #0
+; CHECK-CVT-NEXT:    sshll2  v2.4s, v1.8h, #0
+; CHECK-CVT-NEXT:    sshll   v1.4s, v1.4h, #0
+; CHECK-CVT-NEXT:    scvtf   v2.4s, v2.4s
+; CHECK-CVT-NEXT:    scvtf   v1.4s, v1.4s
+; CHECK-CVT-NEXT:    sshll   v0.8h, v0.8b, #0
+; CHECK-CVT-NEXT:    fcvtn   v2.4h, v2.4s
+; CHECK-CVT-NEXT:    fcvtn   v1.4h, v1.4s
+; CHECK-CVT-NEXT:    mov v1.d[1], v2.d[0]
+; CHECK-CVT-NEXT:    sshll2  v2.4s, v0.8h, #0
+; CHECK-CVT-NEXT:    sshll   v0.4s, v0.4h, #0
+; CHECK-CVT-NEXT:    scvtf   v2.4s, v2.4s
+; CHECK-CVT-NEXT:    scvtf   v0.4s, v0.4s
+; CHECK-CVT-NEXT:    fcvtn   v2.4h, v2.4s
+; CHECK-CVT-NEXT:    fcvtn   v0.4h, v0.4s
+; CHECK-CVT-NEXT:    mov v0.d[1], v2.d[0]
+; CHECK-CVT-NEXT:    ret
 ;
+; CHECK-FP16-LABEL: sitofp_v16i8:
+; CHECK-FP16-NEXT:    sshll2 v1.8h, v0.16b, #0
+; CHECK-FP16-NEXT:    sshll v0.8h, v0.8b, #0
+; CHECK-FP16-NEXT:    scvtf v1.8h, v1.8h
+; CHECK-FP16-NEXT:    scvtf v0.8h, v0.8h
+; CHECK-FP16-NEXT:    ret
   %1 = sitofp <16 x i8> %a to <16 x half>
   ret <16 x half> %1
 }
@@ -391,40 +401,52 @@ define <4 x half> @uitofp_v4i8(<4 x i8> %a) #0 {
 }
 
 define <8 x half> @uitofp_v8i8(<8 x i8> %a) #0 {
-; CHECK-LABEL: uitofp_v8i8:
-; CHECK-NEXT:    ushll v0.8h, v0.8b, #0
-; CHECK-NEXT:    ushll2 v1.4s, v0.8h, #0
-; CHECK-NEXT:    ushll v0.4s, v0.4h, #0
-; CHECK-NEXT:    ucvtf v1.4s, v1.4s
-; CHECK-NEXT:    ucvtf v0.4s, v0.4s
-; CHECK-NEXT:    fcvtn v1.4h, v1.4s
-; CHECK-NEXT:    fcvtn v0.4h, v0.4s
-; CHECK-NEXT:    mov v0.d[1], v1.d[0]
-; CHECK-NEXT:    ret
+; CHECK-CVT-LABEL: uitofp_v8i8:
+; CHECK-CVT-NEXT:    ushll v0.8h, v0.8b, #0
+; CHECK-CVT-NEXT:    ushll2 v1.4s, v0.8h, #0
+; CHECK-CVT-NEXT:    ushll v0.4s, v0.4h, #0
+; CHECK-CVT-NEXT:    ucvtf v1.4s, v1.4s
+; CHECK-CVT-NEXT:    ucvtf v0.4s, v0.4s
+; CHECK-CVT-NEXT:    fcvtn v1.4h, v1.4s
+; CHECK-CVT-NEXT:    fcvtn v0.4h, v0.4s
+; CHECK-CVT-NEXT:    mov v0.d[1], v1.d[0]
+; CHECK-CVT-NEXT:    ret
+;
+; CHECK-FP16-LABEL: uitofp_v8i8:
+; CHECK-FP16-NEXT:    ushll v0.8h, v0.8b, #0
+; CHECK-FP16-NEXT:    ucvtf v0.8h, v0.8h
+; CHECK-FP16-NEXT:    ret
 ;
   %1 = uitofp <8 x i8> %a to <8 x half>
   ret <8 x half> %1
 }
 
 define <16 x half> @uitofp_v16i8(<16 x i8> %a) #0 {
-; CHECK-LABEL: uitofp_v16i8:
-; CHECK-NEXT:    ushll2  v1.8h, v0.16b, #0
-; CHECK-NEXT:    ushll2  v2.4s, v1.8h, #0
-; CHECK-NEXT:    ushll   v1.4s, v1.4h, #0
-; CHECK-NEXT:    ucvtf   v2.4s, v2.4s
-; CHECK-NEXT:    ucvtf   v1.4s, v1.4s
-; CHECK-NEXT:    ushll   v0.8h, v0.8b, #0
-; CHECK-NEXT:    fcvtn   v2.4h, v2.4s
-; CHECK-NEXT:    fcvtn   v1.4h, v1.4s
-; CHECK-NEXT:    mov v1.d[1], v2.d[0]
-; CHECK-NEXT:    ushll2  v2.4s, v0.8h, #0
-; CHECK-NEXT:    ushll   v0.4s, v0.4h, #0
-; CHECK-NEXT:    ucvtf   v2.4s, v2.4s
-; CHECK-NEXT:    ucvtf   v0.4s, v0.4s
-; CHECK-NEXT:    fcvtn   v2.4h, v2.4s
-; CHECK-NEXT:    fcvtn   v0.4h, v0.4s
-; CHECK-NEXT:    mov v0.d[1], v2.d[0]
-; CHECK-NEXT:    ret
+; CHECK-CVT-LABEL: uitofp_v16i8:
+; CHECK-CVT-NEXT:    ushll2  v1.8h, v0.16b, #0
+; CHECK-CVT-NEXT:    ushll2  v2.4s, v1.8h, #0
+; CHECK-CVT-NEXT:    ushll   v1.4s, v1.4h, #0
+; CHECK-CVT-NEXT:    ucvtf   v2.4s, v2.4s
+; CHECK-CVT-NEXT:    ucvtf   v1.4s, v1.4s
+; CHECK-CVT-NEXT:    ushll   v0.8h, v0.8b, #0
+; CHECK-CVT-NEXT:    fcvtn   v2.4h, v2.4s
+; CHECK-CVT-NEXT:    fcvtn   v1.4h, v1.4s
+; CHECK-CVT-NEXT:    mov v1.d[1], v2.d[0]
+; CHECK-CVT-NEXT:    ushll2  v2.4s, v0.8h, #0
+; CHECK-CVT-NEXT:    ushll   v0.4s, v0.4h, #0
+; CHECK-CVT-NEXT:    ucvtf   v2.4s, v2.4s
+; CHECK-CVT-NEXT:    ucvtf   v0.4s, v0.4s
+; CHECK-CVT-NEXT:    fcvtn   v2.4h, v2.4s
+; CHECK-CVT-NEXT:    fcvtn   v0.4h, v0.4s
+; CHECK-CVT-NEXT:    mov v0.d[1], v2.d[0]
+; CHECK-CVT-NEXT:    ret
+;
+; CHECK-FP16-LABEL: uitofp_v16i8:
+; CHECK-FP16-NEXT:    ushll2 v1.8h, v0.16b, #0
+; CHECK-FP16-NEXT:    ushll v0.8h, v0.8b, #0
+; CHECK-FP16-NEXT:    ucvtf v1.8h, v1.8h
+; CHECK-FP16-NEXT:    ucvtf v0.8h, v0.8h
+; CHECK-FP16-NEXT:    ret
 ;
   %1 = uitofp <16 x i8> %a to <16 x half>
   ret <16 x half> %1
