[PATCH] D145184: [AArch64] Prefer to fold dup into fmul/fma as opposed to ld1r

[Dave Green via Phabricator via llvm-commits]

This revision was landed with ongoing or failed builds.
This revision was automatically updated to reflect the committed changes.
Closed by commit rG9aa39481d9eb: [AArch64] Prefer to fold dup into fmul/fma as opposed to ld1r (authored by dmgreen).

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D145184/new/

https://reviews.llvm.org/D145184

Files:
  llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
  llvm/test/CodeGen/AArch64/ld1postmul.ll


Index: llvm/test/CodeGen/AArch64/ld1postmul.ll
===================================================================
--- llvm/test/CodeGen/AArch64/ld1postmul.ll
+++ llvm/test/CodeGen/AArch64/ld1postmul.ll
@@ -63,8 +63,8 @@
 ;
 ; CHECK-FP16-LABEL: fmul_v4f16:
 ; CHECK-FP16:       // %bb.0:
-; CHECK-FP16-NEXT:    ld1r { v1.4h }, [x0], #2
-; CHECK-FP16-NEXT:    fmul v0.4h, v1.4h, v0.4h
+; CHECK-FP16-NEXT:    ldr h1, [x0], #2
+; CHECK-FP16-NEXT:    fmul v0.4h, v0.4h, v1.h[0]
 ; CHECK-FP16-NEXT:    str d0, [x1]
 ; CHECK-FP16-NEXT:    ret
   %l = load half, ptr %p
@@ -93,8 +93,8 @@
 ;
 ; CHECK-FP16-LABEL: fmla_v4f16:
 ; CHECK-FP16:       // %bb.0:
-; CHECK-FP16-NEXT:    ld1r { v2.4h }, [x0], #2
-; CHECK-FP16-NEXT:    fmla v1.4h, v0.4h, v2.4h
+; CHECK-FP16-NEXT:    ldr h2, [x0], #2
+; CHECK-FP16-NEXT:    fmla v1.4h, v0.4h, v2.h[0]
 ; CHECK-FP16-NEXT:    str d1, [x1]
 ; CHECK-FP16-NEXT:    ret
   %l = load half, ptr %p
@@ -110,8 +110,8 @@
 define ptr @fmul_v4f32(ptr %p, ptr %ps, <4 x float> %t) {
 ; CHECK-LABEL: fmul_v4f32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ld1r { v1.4s }, [x0], #4
-; CHECK-NEXT:    fmul v0.4s, v1.4s, v0.4s
+; CHECK-NEXT:    ldr s1, [x0], #4
+; CHECK-NEXT:    fmul v0.4s, v0.4s, v1.s[0]
 ; CHECK-NEXT:    str q0, [x1]
 ; CHECK-NEXT:    ret
   %l = load float, ptr %p
@@ -126,8 +126,8 @@
 define ptr @fmla_v4f32(ptr %p, ptr %ps, <4 x float> %t, <4 x float> %u) {
 ; CHECK-LABEL: fmla_v4f32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ld1r { v2.4s }, [x0], #4
-; CHECK-NEXT:    fmla v1.4s, v0.4s, v2.4s
+; CHECK-NEXT:    ldr s2, [x0], #4
+; CHECK-NEXT:    fmla v1.4s, v0.4s, v2.s[0]
 ; CHECK-NEXT:    str q1, [x1]
 ; CHECK-NEXT:    ret
   %l = load float, ptr %p
@@ -143,8 +143,8 @@
 define ptr @fmul_v2f64(ptr %p, ptr %ps, <2 x double> %t) {
 ; CHECK-LABEL: fmul_v2f64:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ld1r { v1.2d }, [x0], #8
-; CHECK-NEXT:    fmul v0.2d, v1.2d, v0.2d
+; CHECK-NEXT:    ldr d1, [x0], #8
+; CHECK-NEXT:    fmul v0.2d, v0.2d, v1.d[0]
 ; CHECK-NEXT:    str q0, [x1]
 ; CHECK-NEXT:    ret
   %l = load double, ptr %p
@@ -159,8 +159,8 @@
 define ptr @fmla_v2f64(ptr %p, ptr %ps, <2 x double> %t, <2 x double> %u) {
 ; CHECK-LABEL: fmla_v2f64:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ld1r { v2.2d }, [x0], #8
-; CHECK-NEXT:    fmla v1.2d, v0.2d, v2.2d
+; CHECK-NEXT:    ldr d2, [x0], #8
+; CHECK-NEXT:    fmla v1.2d, v0.2d, v2.d[0]
 ; CHECK-NEXT:    str q1, [x1]
 ; CHECK-NEXT:    ret
   %l = load double, ptr %p
Index: llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
===================================================================
--- llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -19348,6 +19348,15 @@
       return SDValue();
   }
 
+  // If there is one use and it can splat the value, prefer that operation.
+  // TODO: This could be expanded to more operations if they reliably use the
+  // index variants.
+  if (N->hasOneUse()) {
+    unsigned UseOpc = N->use_begin()->getOpcode();
+    if (UseOpc == ISD::FMUL || UseOpc == ISD::FMA)
+      return SDValue();
+  }
+
   SDValue Addr = LD->getOperand(1);
   SDValue Vector = N->getOperand(0);
   // Search for a use of the address operand that is an increment.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D145184.503146.patch
Type: text/x-patch
Size: 3242 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230307/13c03236/attachment.bin>