blob: 393f7c47284b04f49dd2cce7bb3cfb02a46942c4 [file]
; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --filter "LV: Found an estimated cost of [0-9]+(.[0-9]+)? for VF 1 For instruction:\s*%v0 = load" --filter "Cost of [0-9]+(.[0-9]+)? for VF [0-9]+: (INTERLEAVE-GROUP with factor [0-9]+,|EXPRESSION|REPLICATE|WIDEN)" --filter "^ ir<.* = load from index" --filter "LV: Selecting VF:" --filter "Estimated cost per lane"
; REQUIRES: asserts
; RUN: opt -mtriple=wasm32 -mattr=+simd128 -passes=loop-vectorize %s --debug-only=loop-vectorize --disable-output < %s 2>&1 | FileCheck %s
; RUN: opt -mtriple=wasm32 -mattr=+simd128 -passes=loop-vectorize %s --vectorizer-maximize-bandwidth --debug-only=loop-vectorize --disable-output < %s 2>&1 | FileCheck %s --check-prefix=MAX
define hidden void @quantize_3d(ptr nofree noundef readonly captures(none) %input,
; CHECK-LABEL: 'quantize_3d'
; CHECK: Cost of 2 for VF 2: WIDEN ir<%value> = load vp<[[VP9:%[0-9]+]]>, align 4 (!alias.scope !3)
; CHECK: Cost of 2 for VF 2: WIDEN ir<%scaled> = fmul fast ir<%value>, ir<%scale>
; CHECK: Cost of 24 for VF 2: WIDEN-INTRINSIC ir<%rounded> = call fast llvm.round(ir<%scaled>)
; CHECK: Cost of 1 for VF 2: WIDEN-CAST ir<%converted> = fptosi ir<%rounded> to i32
; CHECK: Cost of 1 for VF 2: WIDEN-INTRINSIC ir<%max> = call llvm.smax(ir<%converted>, ir<-127>)
; CHECK: Cost of 1 for VF 2: WIDEN-INTRINSIC ir<%min> = call llvm.smin(ir<%max>, ir<127>)
; CHECK: Cost of 4 for VF 2: WIDEN-CAST ir<%truncated> = trunc nsw ir<%min> to i8
; CHECK: Cost of 2 for VF 2: WIDEN store vp<[[VP10:%[0-9]+]]>, ir<%truncated>, align 1 (!alias.scope !4, !noalias !3)
; CHECK: Cost for VF 2: 41 (Estimated cost per lane: 20.5)
; CHECK: Cost of 2 for VF 4: WIDEN ir<%value> = load vp<[[VP9]]>, align 4 (!alias.scope !3)
; CHECK: Cost of 2 for VF 4: WIDEN ir<%scaled> = fmul fast ir<%value>, ir<%scale>
; CHECK: Cost of 48 for VF 4: WIDEN-INTRINSIC ir<%rounded> = call fast llvm.round(ir<%scaled>)
; CHECK: Cost of 1 for VF 4: WIDEN-CAST ir<%converted> = fptosi ir<%rounded> to i32
; CHECK: Cost of 1 for VF 4: WIDEN-INTRINSIC ir<%max> = call llvm.smax(ir<%converted>, ir<-127>)
; CHECK: Cost of 1 for VF 4: WIDEN-INTRINSIC ir<%min> = call llvm.smin(ir<%max>, ir<127>)
; CHECK: Cost of 3 for VF 4: WIDEN-CAST ir<%truncated> = trunc nsw ir<%min> to i8
; CHECK: Cost of 2 for VF 4: WIDEN store vp<[[VP10]]>, ir<%truncated>, align 1 (!alias.scope !4, !noalias !3)
; CHECK: Cost for VF 4: 64 (Estimated cost per lane: 16)
; CHECK: LV: Selecting VF: 4.
;
; MAX-LABEL: 'quantize_3d'
; MAX: Cost of 2 for VF 2: WIDEN ir<%value> = load vp<[[VP9:%[0-9]+]]>, align 4 (!alias.scope !3)
; MAX: Cost of 2 for VF 2: WIDEN ir<%scaled> = fmul fast ir<%value>, ir<%scale>
; MAX: Cost of 24 for VF 2: WIDEN-INTRINSIC ir<%rounded> = call fast llvm.round(ir<%scaled>)
; MAX: Cost of 1 for VF 2: WIDEN-CAST ir<%converted> = fptosi ir<%rounded> to i32
; MAX: Cost of 1 for VF 2: WIDEN-INTRINSIC ir<%max> = call llvm.smax(ir<%converted>, ir<-127>)
; MAX: Cost of 1 for VF 2: WIDEN-INTRINSIC ir<%min> = call llvm.smin(ir<%max>, ir<127>)
; MAX: Cost of 4 for VF 2: WIDEN-CAST ir<%truncated> = trunc nsw ir<%min> to i8
; MAX: Cost of 2 for VF 2: WIDEN store vp<[[VP10:%[0-9]+]]>, ir<%truncated>, align 1 (!alias.scope !4, !noalias !3)
; MAX: Cost for VF 2: 41 (Estimated cost per lane: 20.5)
; MAX: Cost of 2 for VF 4: WIDEN ir<%value> = load vp<[[VP9]]>, align 4 (!alias.scope !3)
; MAX: Cost of 2 for VF 4: WIDEN ir<%scaled> = fmul fast ir<%value>, ir<%scale>
; MAX: Cost of 48 for VF 4: WIDEN-INTRINSIC ir<%rounded> = call fast llvm.round(ir<%scaled>)
; MAX: Cost of 1 for VF 4: WIDEN-CAST ir<%converted> = fptosi ir<%rounded> to i32
; MAX: Cost of 1 for VF 4: WIDEN-INTRINSIC ir<%max> = call llvm.smax(ir<%converted>, ir<-127>)
; MAX: Cost of 1 for VF 4: WIDEN-INTRINSIC ir<%min> = call llvm.smin(ir<%max>, ir<127>)
; MAX: Cost of 3 for VF 4: WIDEN-CAST ir<%truncated> = trunc nsw ir<%min> to i8
; MAX: Cost of 2 for VF 4: WIDEN store vp<[[VP10]]>, ir<%truncated>, align 1 (!alias.scope !4, !noalias !3)
; MAX: Cost for VF 4: 64 (Estimated cost per lane: 16)
; MAX: Cost of 2 for VF 8: WIDEN ir<%value> = load vp<[[VP9]]>, align 4 (!alias.scope !3)
; MAX: Cost of 4 for VF 8: WIDEN ir<%scaled> = fmul fast ir<%value>, ir<%scale>
; MAX: Cost of 96 for VF 8: WIDEN-INTRINSIC ir<%rounded> = call fast llvm.round(ir<%scaled>)
; MAX: Cost of 2 for VF 8: WIDEN-CAST ir<%converted> = fptosi ir<%rounded> to i32
; MAX: Cost of 4 for VF 8: WIDEN-INTRINSIC ir<%max> = call llvm.smax(ir<%converted>, ir<-127>)
; MAX: Cost of 4 for VF 8: WIDEN-INTRINSIC ir<%min> = call llvm.smin(ir<%max>, ir<127>)
; MAX: Cost of 4 for VF 8: WIDEN-CAST ir<%truncated> = trunc nsw ir<%min> to i8
; MAX: Cost of 2 for VF 8: WIDEN store vp<[[VP10]]>, ir<%truncated>, align 1 (!alias.scope !4, !noalias !3)
; MAX: Cost for VF 8: 122 (Estimated cost per lane: 15.3)
; MAX: Cost of 4 for VF 16: WIDEN ir<%value> = load vp<[[VP9]]>, align 4 (!alias.scope !3)
; MAX: Cost of 8 for VF 16: WIDEN ir<%scaled> = fmul fast ir<%value>, ir<%scale>
; MAX: Cost of 192 for VF 16: WIDEN-INTRINSIC ir<%rounded> = call fast llvm.round(ir<%scaled>)
; MAX: Cost of 4 for VF 16: WIDEN-CAST ir<%converted> = fptosi ir<%rounded> to i32
; MAX: Cost of 8 for VF 16: WIDEN-INTRINSIC ir<%max> = call llvm.smax(ir<%converted>, ir<-127>)
; MAX: Cost of 8 for VF 16: WIDEN-INTRINSIC ir<%min> = call llvm.smin(ir<%max>, ir<127>)
; MAX: Cost of 7 for VF 16: WIDEN-CAST ir<%truncated> = trunc nsw ir<%min> to i8
; MAX: Cost of 2 for VF 16: WIDEN store vp<[[VP10]]>, ir<%truncated>, align 1 (!alias.scope !4, !noalias !3)
; MAX: Cost for VF 16: 237 (Estimated cost per lane: 14.8)
; MAX: LV: Selecting VF: 16.
;
ptr nofree noundef writeonly captures(none) %output,
i32 noundef %count,
ptr nofree noundef readonly captures(none) %scale.ptr) {
entry:
%scale = load float, ptr %scale.ptr, align 4
br label %loop
loop:
%iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
%output.ptr = phi ptr [ %output, %entry ], [ %output.next, %loop ]
%input.ptr = phi ptr [ %input, %entry ], [ %input.next, %loop ]
%input.next = getelementptr inbounds nuw i8, ptr %input.ptr, i32 4
%value = load float, ptr %input.ptr, align 4
%scaled = fmul fast float %value, %scale
%rounded = tail call fast noundef nofpclass(nan inf) float @llvm.round.f32(float nofpclass(nan inf) %scaled)
%converted = fptosi float %rounded to i32
%max = tail call i32 @llvm.smax.i32(i32 %converted, i32 -127)
%min = tail call i32 @llvm.smin.i32(i32 %max, i32 127)
%truncated = trunc nsw i32 %min to i8
%output.next = getelementptr inbounds nuw i8, ptr %output.ptr, i32 1
store i8 %truncated, ptr %output.ptr, align 1
%iv.next = add nuw nsw i32 %iv, 1
%exitcond = icmp eq i32 %iv.next, %count
br i1 %exitcond, label %exit, label %loop
exit:
ret void
}
declare float @llvm.round.f32(float)
declare i32 @llvm.smax.i32(i32, i32)
declare i32 @llvm.smin.i32(i32, i32)