[PATCH] D142656: [SVE][codegen] Add pattern for SVE multiply-add accumulate

[Sushant Gokhale via Phabricator via llvm-commits]

sushgokh updated this revision to Diff 510703.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D142656/new/

https://reviews.llvm.org/D142656

Files:
  llvm/lib/Target/AArch64/SVEInstrFormats.td
  llvm/test/CodeGen/AArch64/sve-int-arith.ll


Index: llvm/test/CodeGen/AArch64/sve-int-arith.ll
===================================================================
--- llvm/test/CodeGen/AArch64/sve-int-arith.ll
+++ llvm/test/CodeGen/AArch64/sve-int-arith.ll
@@ -586,8 +586,8 @@
 ; CHECK-LABEL: muladd_i16_positiveAddend:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    ptrue p0.h
-; CHECK-NEXT:    mul z0.h, p0/m, z0.h, z1.h
-; CHECK-NEXT:    add z0.h, z0.h, #255 // =0xff
+; CHECK-NEXT:    mov z2.h, #255 // =0xff
+; CHECK-NEXT:    mad z0.h, p0/m, z1.h, z2.h
 ; CHECK-NEXT:    ret
 {
   %1 = mul <vscale x 8 x i16> %a, %b
@@ -612,8 +612,8 @@
 ; CHECK-LABEL: muladd_i8_positiveAddend:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    ptrue p0.b
-; CHECK-NEXT:    mul z0.b, p0/m, z0.b, z1.b
-; CHECK-NEXT:    add z0.b, z0.b, #15 // =0xf
+; CHECK-NEXT:    mov z2.b, #15 // =0xf
+; CHECK-NEXT:    mad z0.b, p0/m, z1.b, z2.b
 ; CHECK-NEXT:    ret
 {
   %1 = mul <vscale x 16 x i8> %a, %b
@@ -625,8 +625,8 @@
 ; CHECK-LABEL: muladd_i8_negativeAddend:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    ptrue p0.b
-; CHECK-NEXT:    mul z0.b, p0/m, z0.b, z1.b
-; CHECK-NEXT:    add z0.b, z0.b, #241 // =0xf1
+; CHECK-NEXT:    mov z2.b, #-15 // =0xfffffffffffffff1
+; CHECK-NEXT:    mad z0.b, p0/m, z1.b, z2.b
 ; CHECK-NEXT:    ret
 {
   %1 = mul <vscale x 16 x i8> %a, %b
@@ -744,13 +744,14 @@
   ret <vscale x 16 x i8> %2
 }
 
+; TOFIX: Should generate msb for mul+sub in this case. Shuffling operand of sub generates the required msb instruction.
 define <vscale x 8 x i16> @multiple_fused_ops(<vscale x 8 x i16> %a, <vscale x 8 x i16> %b)
 ; CHECK-LABEL: multiple_fused_ops:
 ; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, #200 // =0xc8
 ; CHECK-NEXT:    ptrue p0.h
-; CHECK-NEXT:    movprfx z2, z0
-; CHECK-NEXT:    mul z2.h, p0/m, z2.h, z1.h
-; CHECK-NEXT:    add z2.h, z2.h, #200 // =0xc8
+; CHECK-NEXT:    mov z2.h, w8
+; CHECK-NEXT:    mla z2.h, p0/m, z0.h, z1.h
 ; CHECK-NEXT:    mul z0.h, p0/m, z0.h, z2.h
 ; CHECK-NEXT:    sub z0.h, z0.h, z1.h
 ; CHECK-NEXT:    ret
@@ -771,15 +772,15 @@
 ; CHECK-NEXT:    mov w9, w3
 ; CHECK-NEXT:    mov x8, xzr
 ; CHECK-NEXT:    cntw x10
+; CHECK-NEXT:    mov z0.s, #1 // =0x1
 ; CHECK-NEXT:    ptrue p0.s
 ; CHECK-NEXT:    whilelo p1.s, xzr, x9
 ; CHECK-NEXT:  .LBB70_2: // %vector.body
 ; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    ld1w { z0.s }, p1/z, [x1, x8, lsl #2]
-; CHECK-NEXT:    ld1w { z1.s }, p1/z, [x2, x8, lsl #2]
-; CHECK-NEXT:    mul z0.s, p0/m, z0.s, z1.s
-; CHECK-NEXT:    add z0.s, z0.s, #1 // =0x1
-; CHECK-NEXT:    st1w { z0.s }, p1, [x0, x8, lsl #2]
+; CHECK-NEXT:    ld1w { z1.s }, p1/z, [x1, x8, lsl #2]
+; CHECK-NEXT:    ld1w { z2.s }, p1/z, [x2, x8, lsl #2]
+; CHECK-NEXT:    mad z1.s, p0/m, z2.s, z0.s
+; CHECK-NEXT:    st1w { z1.s }, p1, [x0, x8, lsl #2]
 ; CHECK-NEXT:    add x8, x8, x10
 ; CHECK-NEXT:    whilelo p1.s, x8, x9
 ; CHECK-NEXT:    b.mi .LBB70_2
Index: llvm/lib/Target/AArch64/SVEInstrFormats.td
===================================================================
--- llvm/lib/Target/AArch64/SVEInstrFormats.td
+++ llvm/lib/Target/AArch64/SVEInstrFormats.td
@@ -3198,10 +3198,12 @@
   def _UNDEF_S : PredThreeOpPseudo<NAME # _S, ZPR32, FalseLanesUndef>;
   def _UNDEF_D : PredThreeOpPseudo<NAME # _D, ZPR64, FalseLanesUndef>;
 
-  def : SVE_4_Op_Pat<nxv16i8, op, nxv16i1, nxv16i8, nxv16i8, nxv16i8, !cast<Instruction>(NAME # _UNDEF_B)>;
-  def : SVE_4_Op_Pat<nxv8i16, op, nxv8i1,  nxv8i16, nxv8i16, nxv8i16, !cast<Instruction>(NAME # _UNDEF_H)>;
-  def : SVE_4_Op_Pat<nxv4i32, op, nxv4i1,  nxv4i32, nxv4i32, nxv4i32, !cast<Instruction>(NAME # _UNDEF_S)>;
-  def : SVE_4_Op_Pat<nxv2i64, op, nxv2i1,  nxv2i64, nxv2i64, nxv2i64, !cast<Instruction>(NAME # _UNDEF_D)>;
+  let  AddedComplexity = 9 in {
+    def : SVE_4_Op_Pat<nxv16i8, op, nxv16i1, nxv16i8, nxv16i8, nxv16i8, !cast<Instruction>(NAME # _UNDEF_B)>;
+    def : SVE_4_Op_Pat<nxv8i16, op, nxv8i1,  nxv8i16, nxv8i16, nxv8i16, !cast<Instruction>(NAME # _UNDEF_H)>;
+    def : SVE_4_Op_Pat<nxv4i32, op, nxv4i1,  nxv4i32, nxv4i32, nxv4i32, !cast<Instruction>(NAME # _UNDEF_S)>;
+    def : SVE_4_Op_Pat<nxv2i64, op, nxv2i1,  nxv2i64, nxv2i64, nxv2i64, !cast<Instruction>(NAME # _UNDEF_D)>;
+  }
 }
 
 //===----------------------------------------------------------------------===//


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D142656.510703.patch
Type: text/x-patch
Size: 4284 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230404/4f7565d1/attachment.bin>