b5fb26951a
(a[0] + a[1] + a[2] + a[3]) - (b[0] + b[1] + b[2] +b[3]) --> (a[0] - b[0]) + (a[1] - b[1]) + (a[2] - b[2]) + (a[3] - b[3]) This should be the last step in solving PR43953: https://bugs.llvm.org/show_bug.cgi?id=43953 We started emitting reduction intrinsics with: D80867/ rGe50059f6b6b3 So it's a relatively easy pattern match now to re-order those ops. Also, I have not seen any complaints for the switch to intrinsics yet, so I'll propose to remove the "experimental" tag from the intrinsics soon. Differential Revision: https://reviews.llvm.org/D81491
84 lines
4.4 KiB
LLVM
84 lines
4.4 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
|
|
; RUN: opt < %s -instcombine -S | FileCheck %s
|
|
|
|
declare float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float, <4 x float>)
|
|
declare float @llvm.experimental.vector.reduce.v2.fadd.f32.v8f32(float, <8 x float>)
|
|
declare void @use_f32(float)
|
|
|
|
define float @diff_of_sums_v4f32(float %a0, <4 x float> %v0, float %a1, <4 x float> %v1) {
|
|
; CHECK-LABEL: @diff_of_sums_v4f32(
|
|
; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz <4 x float> [[V0:%.*]], [[V1:%.*]]
|
|
; CHECK-NEXT: [[TMP2:%.*]] = call reassoc nsz float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float [[A0:%.*]], <4 x float> [[TMP1]])
|
|
; CHECK-NEXT: [[R:%.*]] = fsub reassoc nsz float [[TMP2]], [[A1:%.*]]
|
|
; CHECK-NEXT: ret float [[R]]
|
|
;
|
|
%r0 = call float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %a0, <4 x float> %v0)
|
|
%r1 = call float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %a1, <4 x float> %v1)
|
|
%r = fsub reassoc nsz float %r0, %r1
|
|
ret float %r
|
|
}
|
|
|
|
; negative test - fsub must allow reassociation
|
|
|
|
define float @diff_of_sums_v4f32_fmf(float %a0, <4 x float> %v0, float %a1, <4 x float> %v1) {
|
|
; CHECK-LABEL: @diff_of_sums_v4f32_fmf(
|
|
; CHECK-NEXT: [[R0:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float [[A0:%.*]], <4 x float> [[V0:%.*]])
|
|
; CHECK-NEXT: [[R1:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float [[A1:%.*]], <4 x float> [[V1:%.*]])
|
|
; CHECK-NEXT: [[R:%.*]] = fsub nnan ninf nsz float [[R0]], [[R1]]
|
|
; CHECK-NEXT: ret float [[R]]
|
|
;
|
|
%r0 = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %a0, <4 x float> %v0)
|
|
%r1 = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %a1, <4 x float> %v1)
|
|
%r = fsub ninf nnan nsz float %r0, %r1
|
|
ret float %r
|
|
}
|
|
|
|
; negative test - extra uses could create extra instructions
|
|
|
|
define float @diff_of_sums_extra_use1(float %a0, <4 x float> %v0, float %a1, <4 x float> %v1) {
|
|
; CHECK-LABEL: @diff_of_sums_extra_use1(
|
|
; CHECK-NEXT: [[R0:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float [[A0:%.*]], <4 x float> [[V0:%.*]])
|
|
; CHECK-NEXT: call void @use_f32(float [[R0]])
|
|
; CHECK-NEXT: [[R1:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float [[A1:%.*]], <4 x float> [[V1:%.*]])
|
|
; CHECK-NEXT: [[R:%.*]] = fsub fast float [[R0]], [[R1]]
|
|
; CHECK-NEXT: ret float [[R]]
|
|
;
|
|
%r0 = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %a0, <4 x float> %v0)
|
|
call void @use_f32(float %r0)
|
|
%r1 = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %a1, <4 x float> %v1)
|
|
%r = fsub fast float %r0, %r1
|
|
ret float %r
|
|
}
|
|
|
|
; negative test - extra uses could create extra instructions
|
|
|
|
define float @diff_of_sums_extra_use2(float %a0, <4 x float> %v0, float %a1, <4 x float> %v1) {
|
|
; CHECK-LABEL: @diff_of_sums_extra_use2(
|
|
; CHECK-NEXT: [[R0:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float [[A0:%.*]], <4 x float> [[V0:%.*]])
|
|
; CHECK-NEXT: [[R1:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float [[A1:%.*]], <4 x float> [[V1:%.*]])
|
|
; CHECK-NEXT: call void @use_f32(float [[R1]])
|
|
; CHECK-NEXT: [[R:%.*]] = fsub fast float [[R0]], [[R1]]
|
|
; CHECK-NEXT: ret float [[R]]
|
|
;
|
|
%r0 = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %a0, <4 x float> %v0)
|
|
%r1 = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %a1, <4 x float> %v1)
|
|
call void @use_f32(float %r1)
|
|
%r = fsub fast float %r0, %r1
|
|
ret float %r
|
|
}
|
|
|
|
; negative test - can't reassociate different vector types
|
|
|
|
define float @diff_of_sums_type_mismatch(float %a0, <4 x float> %v0, float %a1, <8 x float> %v1) {
|
|
; CHECK-LABEL: @diff_of_sums_type_mismatch(
|
|
; CHECK-NEXT: [[R0:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float [[A0:%.*]], <4 x float> [[V0:%.*]])
|
|
; CHECK-NEXT: [[R1:%.*]] = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v8f32(float [[A1:%.*]], <8 x float> [[V1:%.*]])
|
|
; CHECK-NEXT: [[R:%.*]] = fsub fast float [[R0]], [[R1]]
|
|
; CHECK-NEXT: ret float [[R]]
|
|
;
|
|
%r0 = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %a0, <4 x float> %v0)
|
|
%r1 = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v8f32(float %a1, <8 x float> %v1)
|
|
%r = fsub fast float %r0, %r1
|
|
ret float %r
|
|
}
|