blob: 4d12886503d01e704621d23072df242c0a29df87 [file] [edit]
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --filter "(:|reduce\.fadd)" --version 6
; Two targets exercise the two halves of the fp_unsafe hint.
;
; ARMv7 NEON is neither AArch64 nor MVE, so its vector FP units flush
; subnormals to zero and TTI::isFPVectorizationPotentiallyUnsafe() is true
; there. The hint controls the resulting isPotentiallyUnsafe() guard.
;
; Armv8.1-M MVE has IEEE-754 compliant vector FP, so that guard is inactive
; and the hint only affects FP reordering.
; RUN: opt -passes='loop-vectorize' -mtriple=armv7-none-linux-gnueabihf -mattr=+neon \
; RUN: -force-vector-width=4 -S < %s | FileCheck %s --check-prefix=NEON
; RUN: opt -passes='loop-vectorize' -mtriple=thumbv8.1m.main-none-eabi -mattr=+mve.fp \
; RUN: -force-vector-width=4 -S < %s | FileCheck %s --check-prefix=MVE
; No hint: the guard rejects the loop on NEON; on MVE it is inactive, so the
; loop vectorizes with a reordered reduction.
; UTC_ARGS: --disable
define float @no_hint(ptr %A, i64 %n) {
; NEON-LABEL: define float @no_hint(
; NEON-SAME: ptr [[A:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
; NEON: [[ENTRY:.*:]]
; NEON: [[LOOP:.*:]]
; NEON: [[EXIT:.*:]]
; NEON-NOT: vector.body
;
; MVE-LABEL: define float @no_hint(
; MVE-SAME: ptr [[A:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
; MVE: [[ENTRY:.*:]]
; MVE: [[VECTOR_PH:.*:]]
; MVE: [[VECTOR_BODY:.*:]]
; MVE: [[MIDDLE_BLOCK:.*:]]
; MVE: [[TMP4:%.*]] = call float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[TMP2:%.*]])
; MVE: [[SCALAR_PH:.*:]]
; MVE: [[LOOP:.*:]]
; MVE: [[EXIT:.*:]]
;
entry:
br label %loop
loop:
%i = phi i64 [ 0, %entry ], [ %i.next, %loop ]
%sum = phi float [ 0.0, %entry ], [ %sum.next, %loop ]
%ptr = getelementptr inbounds float, ptr %A, i64 %i
%val = load float, ptr %ptr, align 4
%sum.next = fadd float %sum, %val
%i.next = add nuw nsw i64 %i, 1
%cond = icmp eq i64 %i.next, %n
br i1 %cond, label %exit, label %loop
exit:
ret float %sum.next
}
; UTC_ARGS: --enable
; fp_unsafe.enable: suppresses the guard, so the loop vectorizes on both
; targets.
define float @fp_unsafe_enable(ptr %A, i64 %n) {
; NEON-LABEL: define float @fp_unsafe_enable(
; NEON-SAME: ptr [[A:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
; NEON: [[ENTRY:.*:]]
; NEON: [[VECTOR_PH:.*:]]
; NEON: [[VECTOR_BODY:.*:]]
; NEON: [[MIDDLE_BLOCK:.*:]]
; NEON: [[TMP4:%.*]] = call float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[TMP2:%.*]])
; NEON: [[SCALAR_PH:.*:]]
; NEON: [[LOOP:.*:]]
; NEON: [[EXIT:.*:]]
;
; MVE-LABEL: define float @fp_unsafe_enable(
; MVE-SAME: ptr [[A:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
; MVE: [[ENTRY:.*:]]
; MVE: [[VECTOR_PH:.*:]]
; MVE: [[VECTOR_BODY:.*:]]
; MVE: [[MIDDLE_BLOCK:.*:]]
; MVE: [[TMP4:%.*]] = call float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[TMP2:%.*]])
; MVE: [[SCALAR_PH:.*:]]
; MVE: [[LOOP:.*:]]
; MVE: [[EXIT:.*:]]
;
entry:
br label %loop
loop:
%i = phi i64 [ 0, %entry ], [ %i.next, %loop ]
%sum = phi float [ 0.0, %entry ], [ %sum.next, %loop ]
%ptr = getelementptr inbounds float, ptr %A, i64 %i
%val = load float, ptr %ptr, align 4
%sum.next = fadd float %sum, %val
%i.next = add nuw nsw i64 %i, 1
%cond = icmp eq i64 %i.next, %n
br i1 %cond, label %exit, label %loop, !llvm.loop !0
exit:
ret float %sum.next
}
; fp_unsafe.disable: rejected on both targets, but for different reasons. On
; NEON the guard rejects it (UnsafeFP); on MVE the guard is inactive and the
; rejection comes from suppressed FP reordering (CantReorderFPOps).
; UTC_ARGS: --disable
define float @fp_unsafe_disable(ptr %A, i64 %n) {
; NEON-LABEL: define float @fp_unsafe_disable(
; NEON-SAME: ptr [[A:%.*]], i64 [[N:%.*]]) #[[ATTR0]] {
; NEON: [[ENTRY:.*:]]
; NEON: [[LOOP:.*:]]
; NEON: [[EXIT:.*:]]
; NEON-NOT: vector.body
;
; MVE-LABEL: define float @fp_unsafe_disable(
; MVE-SAME: ptr [[A:%.*]], i64 [[N:%.*]]) #[[ATTR0]] {
; MVE: [[ENTRY:.*:]]
; MVE: [[LOOP:.*:]]
; MVE: [[EXIT:.*:]]
; MVE-NOT: vector.body
;
entry:
br label %loop
loop:
%i = phi i64 [ 0, %entry ], [ %i.next, %loop ]
%sum = phi float [ 0.0, %entry ], [ %sum.next, %loop ]
%ptr = getelementptr inbounds float, ptr %A, i64 %i
%val = load float, ptr %ptr, align 4
%sum.next = fadd float %sum, %val
%i.next = add nuw nsw i64 %i, 1
%cond = icmp eq i64 %i.next, %n
br i1 %cond, label %exit, label %loop, !llvm.loop !2
exit:
ret float %sum.next
}
!0 = distinct !{!0, !1}
!1 = !{!"llvm.loop.vectorize.fp_unsafe.enable"}
!2 = distinct !{!2, !3}
!3 = !{!"llvm.loop.vectorize.fp_unsafe.disable"}