[WebAssembly] Expand vector frem instructions (#174854)
Commit
6ad41bcc49
changed how frem is expanded during legalization and it
broke WebAssembly but we were missing test coverage. We want to maintain
our previous behavior of unrolling vectors and using a libcall to
implement scalar frem. I'm not sure why this now has to be different
(in ISelLowering) from other libcalls like fsin which work the same way
in the end, but this code does accurately describe what we want.
Fixes: https://github.com/emscripten-core/emscripten/issues/25991
This commit is contained in:
Derek Schuff
GitHub
parent
208687e4a3
commit
7a22bea512
3 changed files with 87 additions and 60 deletions
|
|
@ -139,7 +139,11 @@ WebAssemblyTargetLowering::WebAssemblyTargetLowering(
|
|||
// Expand floating-point library function operators.
|
||||
for (auto Op : {ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FMA})
|
||||
setOperationAction(Op, T, Expand);
|
||||
setOperationAction(ISD::FREM, T, LibCall);
|
||||
// Expand vector FREM, but use a libcall rather than an expansion for scalar
|
||||
if (MVT(T).isVector())
|
||||
setOperationAction(ISD::FREM, T, Expand);
|
||||
else
|
||||
setOperationAction(ISD::FREM, T, LibCall);
|
||||
// Note supported floating-point library function operators that otherwise
|
||||
// default to expand.
|
||||
for (auto Op : {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT,
|
||||
|
|
|
|||
|
|
@ -32,71 +32,77 @@ define fp128 @fp128libcalls(fp128 %x, fp128 %y, i32 %z) {
|
|||
; CHECK-LABEL: fp128libcalls:
|
||||
; CHECK: .functype fp128libcalls (i32, i64, i64, i64, i64, i32) -> ()
|
||||
; CHECK-NEXT: # %bb.0:
|
||||
; CHECK-NEXT: global.get $push18=, __stack_pointer
|
||||
; CHECK-NEXT: i32.const $push19=, 144
|
||||
; CHECK-NEXT: i32.sub $push39=, $pop18, $pop19
|
||||
; CHECK-NEXT: local.tee $push38=, $6=, $pop39
|
||||
; CHECK-NEXT: global.set __stack_pointer, $pop38
|
||||
; CHECK-NEXT: i32.const $push36=, 128
|
||||
; CHECK-NEXT: global.get $push20=, __stack_pointer
|
||||
; CHECK-NEXT: i32.const $push21=, 160
|
||||
; CHECK-NEXT: i32.sub $push43=, $pop20, $pop21
|
||||
; CHECK-NEXT: local.tee $push42=, $6=, $pop43
|
||||
; CHECK-NEXT: global.set __stack_pointer, $pop42
|
||||
; CHECK-NEXT: i32.const $push40=, 144
|
||||
; CHECK-NEXT: i32.add $push41=, $6, $pop40
|
||||
; CHECK-NEXT: call __addtf3, $pop41, $1, $2, $3, $4
|
||||
; CHECK-NEXT: i32.const $push38=, 128
|
||||
; CHECK-NEXT: i32.add $push39=, $6, $pop38
|
||||
; CHECK-NEXT: i64.load $push1=, 144($6)
|
||||
; CHECK-NEXT: i64.load $push0=, 152($6)
|
||||
; CHECK-NEXT: call __multf3, $pop39, $pop1, $pop0, $3, $4
|
||||
; CHECK-NEXT: i32.const $push36=, 112
|
||||
; CHECK-NEXT: i32.add $push37=, $6, $pop36
|
||||
; CHECK-NEXT: call __addtf3, $pop37, $1, $2, $3, $4
|
||||
; CHECK-NEXT: i32.const $push34=, 112
|
||||
; CHECK-NEXT: i64.load $push3=, 128($6)
|
||||
; CHECK-NEXT: i64.load $push2=, 136($6)
|
||||
; CHECK-NEXT: call __divtf3, $pop37, $pop3, $pop2, $3, $4
|
||||
; CHECK-NEXT: i32.const $push34=, 96
|
||||
; CHECK-NEXT: i32.add $push35=, $6, $pop34
|
||||
; CHECK-NEXT: i64.load $push1=, 128($6)
|
||||
; CHECK-NEXT: i64.load $push0=, 136($6)
|
||||
; CHECK-NEXT: call __multf3, $pop35, $pop1, $pop0, $3, $4
|
||||
; CHECK-NEXT: i32.const $push32=, 96
|
||||
; CHECK-NEXT: i64.load $push5=, 112($6)
|
||||
; CHECK-NEXT: i64.load $push4=, 120($6)
|
||||
; CHECK-NEXT: call fmodl, $pop35, $pop5, $pop4, $3, $4
|
||||
; CHECK-NEXT: i32.const $push32=, 80
|
||||
; CHECK-NEXT: i32.add $push33=, $6, $pop32
|
||||
; CHECK-NEXT: i64.load $push3=, 112($6)
|
||||
; CHECK-NEXT: i64.load $push2=, 120($6)
|
||||
; CHECK-NEXT: call __divtf3, $pop33, $pop3, $pop2, $3, $4
|
||||
; CHECK-NEXT: i32.const $push30=, 80
|
||||
; CHECK-NEXT: i64.load $push7=, 96($6)
|
||||
; CHECK-NEXT: i64.load $push6=, 104($6)
|
||||
; CHECK-NEXT: call sqrtl, $pop33, $pop7, $pop6
|
||||
; CHECK-NEXT: i32.const $push30=, 64
|
||||
; CHECK-NEXT: i32.add $push31=, $6, $pop30
|
||||
; CHECK-NEXT: i64.load $push5=, 96($6)
|
||||
; CHECK-NEXT: i64.load $push4=, 104($6)
|
||||
; CHECK-NEXT: call sqrtl, $pop31, $pop5, $pop4
|
||||
; CHECK-NEXT: i32.const $push28=, 64
|
||||
; CHECK-NEXT: i64.load $push9=, 80($6)
|
||||
; CHECK-NEXT: i64.load $push8=, 88($6)
|
||||
; CHECK-NEXT: call floorl, $pop31, $pop9, $pop8
|
||||
; CHECK-NEXT: i32.const $push28=, 48
|
||||
; CHECK-NEXT: i32.add $push29=, $6, $pop28
|
||||
; CHECK-NEXT: i64.load $push7=, 80($6)
|
||||
; CHECK-NEXT: i64.load $push6=, 88($6)
|
||||
; CHECK-NEXT: call floorl, $pop29, $pop7, $pop6
|
||||
; CHECK-NEXT: i32.const $push26=, 48
|
||||
; CHECK-NEXT: i64.load $push11=, 64($6)
|
||||
; CHECK-NEXT: i64.load $push10=, 72($6)
|
||||
; CHECK-NEXT: call powl, $pop29, $pop11, $pop10, $3, $4
|
||||
; CHECK-NEXT: i32.const $push26=, 32
|
||||
; CHECK-NEXT: i32.add $push27=, $6, $pop26
|
||||
; CHECK-NEXT: i64.load $push9=, 64($6)
|
||||
; CHECK-NEXT: i64.load $push8=, 72($6)
|
||||
; CHECK-NEXT: call powl, $pop27, $pop9, $pop8, $3, $4
|
||||
; CHECK-NEXT: i32.const $push24=, 32
|
||||
; CHECK-NEXT: i64.load $push13=, 48($6)
|
||||
; CHECK-NEXT: i64.load $push12=, 56($6)
|
||||
; CHECK-NEXT: call __powitf2, $pop27, $pop13, $pop12, $5
|
||||
; CHECK-NEXT: i32.const $push24=, 16
|
||||
; CHECK-NEXT: i32.add $push25=, $6, $pop24
|
||||
; CHECK-NEXT: i64.load $push11=, 48($6)
|
||||
; CHECK-NEXT: i64.load $push10=, 56($6)
|
||||
; CHECK-NEXT: call __powitf2, $pop25, $pop11, $pop10, $5
|
||||
; CHECK-NEXT: i32.const $push22=, 16
|
||||
; CHECK-NEXT: i64.load $push15=, 32($6)
|
||||
; CHECK-NEXT: i64.load $push14=, 40($6)
|
||||
; CHECK-NEXT: call truncl, $pop25, $pop15, $pop14
|
||||
; CHECK-NEXT: i64.load $push17=, 16($6)
|
||||
; CHECK-NEXT: i64.load $push16=, 24($6)
|
||||
; CHECK-NEXT: call nearbyintl, $6, $pop17, $pop16
|
||||
; CHECK-NEXT: i64.load $push18=, 8($6)
|
||||
; CHECK-NEXT: i64.store 8($0), $pop18
|
||||
; CHECK-NEXT: i64.load $push19=, 0($6)
|
||||
; CHECK-NEXT: i64.store 0($0), $pop19
|
||||
; CHECK-NEXT: i32.const $push22=, 160
|
||||
; CHECK-NEXT: i32.add $push23=, $6, $pop22
|
||||
; CHECK-NEXT: i64.load $push13=, 32($6)
|
||||
; CHECK-NEXT: i64.load $push12=, 40($6)
|
||||
; CHECK-NEXT: call truncl, $pop23, $pop13, $pop12
|
||||
; CHECK-NEXT: i64.load $push15=, 16($6)
|
||||
; CHECK-NEXT: i64.load $push14=, 24($6)
|
||||
; CHECK-NEXT: call nearbyintl, $6, $pop15, $pop14
|
||||
; CHECK-NEXT: i64.load $push16=, 8($6)
|
||||
; CHECK-NEXT: i64.store 8($0), $pop16
|
||||
; CHECK-NEXT: i64.load $push17=, 0($6)
|
||||
; CHECK-NEXT: i64.store 0($0), $pop17
|
||||
; CHECK-NEXT: i32.const $push20=, 144
|
||||
; CHECK-NEXT: i32.add $push21=, $6, $pop20
|
||||
; CHECK-NEXT: global.set __stack_pointer, $pop21
|
||||
; CHECK-NEXT: global.set __stack_pointer, $pop23
|
||||
; CHECK-NEXT: return
|
||||
%a = fadd fp128 %x, %y
|
||||
%b = fmul fp128 %a, %y
|
||||
%c = fdiv fp128 %b, %y
|
||||
%d = frem fp128 %c, %y
|
||||
; libm calls
|
||||
%d = call fp128 @llvm.sqrt.f128(fp128 %c)
|
||||
%e = call fp128 @llvm.floor.f128(fp128 %d)
|
||||
%f = call fp128 @llvm.pow.f128(fp128 %e, fp128 %y)
|
||||
%g = call fp128 @llvm.powi.f128.i32(fp128 %f, i32 %z)
|
||||
%h = call fp128 @llvm.trunc.f128(fp128 %g)
|
||||
%i = call fp128 @llvm.nearbyint.f128(fp128 %h)
|
||||
ret fp128 %i
|
||||
%e = call fp128 @llvm.sqrt.f128(fp128 %d)
|
||||
%f = call fp128 @llvm.floor.f128(fp128 %e)
|
||||
%g = call fp128 @llvm.pow.f128(fp128 %f, fp128 %y)
|
||||
%h = call fp128 @llvm.powi.f128.i32(fp128 %g, i32 %z)
|
||||
%i = call fp128 @llvm.trunc.f128(fp128 %h)
|
||||
%j = call fp128 @llvm.nearbyint.f128(fp128 %i)
|
||||
ret fp128 %j
|
||||
}
|
||||
|
||||
define i128 @i128libcalls(i128 %x, i128 %y) {
|
||||
|
|
@ -141,9 +147,9 @@ define double @f64libcalls(double %x, double %y, i32 %z) {
|
|||
; CHECK-NEXT: # %bb.0:
|
||||
; CHECK-NEXT: global.get $push12=, __stack_pointer
|
||||
; CHECK-NEXT: i32.const $push13=, 16
|
||||
; CHECK-NEXT: i32.sub $push21=, $pop12, $pop13
|
||||
; CHECK-NEXT: local.tee $push20=, $3=, $pop21
|
||||
; CHECK-NEXT: global.set __stack_pointer, $pop20
|
||||
; CHECK-NEXT: i32.sub $push23=, $pop12, $pop13
|
||||
; CHECK-NEXT: local.tee $push22=, $3=, $pop23
|
||||
; CHECK-NEXT: global.set __stack_pointer, $pop22
|
||||
; CHECK-NEXT: call $push0=, tan, $0
|
||||
; CHECK-NEXT: call $push1=, cos, $pop0
|
||||
; CHECK-NEXT: call $push2=, log10, $pop1
|
||||
|
|
@ -152,9 +158,11 @@ define double @f64libcalls(double %x, double %y, i32 %z) {
|
|||
; CHECK-NEXT: call $push5=, log, $pop4
|
||||
; CHECK-NEXT: call $push6=, exp, $pop5
|
||||
; CHECK-NEXT: call $push7=, exp10, $pop6
|
||||
; CHECK-NEXT: call $push8=, cbrt, $pop7
|
||||
; CHECK-NEXT: call $push9=, lround, $pop8
|
||||
; CHECK-NEXT: call $push10=, ldexp, $0, $pop9
|
||||
; CHECK-NEXT: call $push21=, cbrt, $pop7
|
||||
; CHECK-NEXT: local.tee $push20=, $1=, $pop21
|
||||
; CHECK-NEXT: call $push8=, lround, $pop20
|
||||
; CHECK-NEXT: call $push9=, ldexp, $0, $pop8
|
||||
; CHECK-NEXT: call $push10=, fmod, $pop9, $1
|
||||
; CHECK-NEXT: i32.const $push18=, 4
|
||||
; CHECK-NEXT: i32.add $push19=, $3, $pop18
|
||||
; CHECK-NEXT: call $0=, frexp, $pop10, $pop19
|
||||
|
|
@ -180,7 +188,8 @@ define double @f64libcalls(double %x, double %y, i32 %z) {
|
|||
%h = call fast double @llvm.pow.f64(double %g, double 0x3FD5555555555555)
|
||||
%i = call i32 @llvm.lround(double %h)
|
||||
%j = call double @llvm.ldexp.f64.i32(double %x, i32 %i);
|
||||
%result = call {double, i32} @llvm.frexp.f64.i32(double %j)
|
||||
%l = frem double %j, %h;
|
||||
%result = call {double, i32} @llvm.frexp.f64.i32(double %l)
|
||||
%result.0 = extractvalue { double, i32 } %result, 0
|
||||
%result.1 = extractvalue { double, i32 } %result, 1
|
||||
%resultModf = call {double, double} @llvm.modf.f64(double %result.0)
|
||||
|
|
|
|||
|
|
@ -353,6 +353,13 @@ define <2 x i64> @rotr_v2i64(<2 x i64> %x, <2 x i64> %y) {
|
|||
; 4 x f32
|
||||
; ==============================================================================
|
||||
|
||||
; CHECK-LABEL: frem_v4f32:
|
||||
; CHECK: call $push[[L:[0-9]+]]=, fmod
|
||||
define <4 x float> @frem_v4f32(<4 x float> %a, <4 x float> %b) #0 {
|
||||
%frem = frem <4 x float> %a, %b
|
||||
ret <4 x float> %frem
|
||||
}
|
||||
|
||||
; CHECK-LABEL: copysign_v4f32:
|
||||
; CHECK: v128.const
|
||||
; CHECK-NEXT: v128.and
|
||||
|
|
@ -457,6 +464,13 @@ define <4 x float> @round_v4f32(<4 x float> %x) {
|
|||
; 2 x f64
|
||||
; ==============================================================================
|
||||
|
||||
; CHECK-LABEL: frem_v2f64:
|
||||
; CHECK: call $push[[L:[0-9]+]]=, fmod
|
||||
define <2 x double> @frem_v2f64(<2 x double> %a, <2 x double> %b) #0 {
|
||||
%frem = frem <2 x double> %a, %b
|
||||
ret <2 x double> %frem
|
||||
}
|
||||
|
||||
; CHECK-LABEL: copysign_v2f64:
|
||||
; CHECK: v128.const
|
||||
; CHECK-NEXT: v128.and
|
||||
|
|
|
|||
Loading…
Reference in a new issue