Skip to content

A lightweight WebAssembly runtime that is fast, secure, and standards-compliant

x64: Lower fcvt_to_{u,s}int{,_sat} to i8 and i16 (#14778)

Ilyas Mallah2d9c19b0fe7f
diff --git a/cranelift/codegen/src/isa/x64/lower.isle b/cranelift/codegen/src/isa/x64/lower.isle
index 453cbe654d..7940f326ab 100644
--- a/cranelift/codegen/src/isa/x64/lower.isle
+++ b/cranelift/codegen/src/isa/x64/lower.isle
@@ -3889,6 +3889,62 @@
(rule (lower (fcvt_to_sint_sat (fits_in_64 out_ty) val @ (value_type (ty_scalar_float _))))
(cvt_float_to_sint_seq out_ty val true))
+;; `i8` and `i16` results reuse the `i32` conversion and then clamp
+;; (saturating) or range check (trapping) its result.
+(rule 2 (lower (fcvt_to_uint (ty_8_or_16 out_ty) val @ (value_type (ty_scalar_float _))))
+ (trap_if_outside_i32 (cvt_float_to_sint_seq $I32 val false)
+ (narrow_int_min out_ty false)
+ (narrow_int_max out_ty false)))
+
+(rule 2 (lower (fcvt_to_uint_sat (ty_8_or_16 out_ty) val @ (value_type (ty_scalar_float _))))
+ (clamp_i32 (cvt_float_to_sint_seq $I32 val true)
+ (narrow_int_min out_ty false)
+ (narrow_int_max out_ty false)))
+
+(rule 2 (lower (fcvt_to_sint (ty_8_or_16 out_ty) val @ (value_type (ty_scalar_float _))))
+ (trap_if_outside_i32 (cvt_float_to_sint_seq $I32 val false)
+ (narrow_int_min out_ty true)
+ (narrow_int_max out_ty true)))
+
+(rule 2 (lower (fcvt_to_sint_sat (ty_8_or_16 out_ty) val @ (value_type (ty_scalar_float _))))
+ (clamp_i32 (cvt_float_to_sint_seq $I32 val true)
+ (narrow_int_min out_ty true)
+ (narrow_int_max out_ty true)))
+
+;; Range of an `i8` or `i16` result as `i32` immediates. The `bool` is whether
+;; the result is signed.
+(decl narrow_int_min (Type bool) u32)
+(rule (narrow_int_min $I8 true) 0xffffff80)
+(rule (narrow_int_min $I16 true) 0xffff8000)
+(rule (narrow_int_min _ false) 0)
+
+(decl narrow_int_max (Type bool) u32)
+(rule (narrow_int_max $I8 true) 0x7f)
+(rule (narrow_int_max $I16 true) 0x7fff)
+(rule (narrow_int_max $I8 false) 0xff)
+(rule (narrow_int_max $I16 false) 0xffff)
+
+;; Clamps the signed `i32` in `x` to `[min, max]`.
+(decl clamp_i32 (Gpr u32 u32) Gpr)
+(rule (clamp_i32 x min max)
+ (let ((max_reg Gpr (imm $I32 (u32_into_u64 max)))
+ (x Gpr (with_flags_reg (x64_cmp $I32 x (RegMemImm.Imm max))
+ (cmove $I32 (CC.NLE) max_reg x)))
+ (min_reg Gpr (imm $I32 (u32_into_u64 min))))
+ (with_flags_reg (x64_cmp $I32 x (RegMemImm.Imm min))
+ (cmove $I32 (CC.L) min_reg x))))
+
+;; Traps with `int_ovf` unless the signed `i32` in `x` is in `[min, max]`.
+(decl trap_if_outside_i32 (Gpr u32 u32) Gpr)
+(rule (trap_if_outside_i32 x min max)
+ (let ((_ InstOutput (side_effect (with_flags_side_effect
+ (x64_cmp $I32 x (RegMemImm.Imm max))
+ (trap_if (CC.NLE) (TrapCode.INTEGER_OVERFLOW)))))
+ (_ InstOutput (side_effect (with_flags_side_effect
+ (x64_cmp $I32 x (RegMemImm.Imm min))
+ (trap_if (CC.L) (TrapCode.INTEGER_OVERFLOW))))))
+ x))
+
;; The x64 backend currently only supports these two type combinations.
(rule 1 (lower (fcvt_to_sint_sat $I32X4 val @ (value_type $F32X4)))
(let ((src Xmm val)
diff --git a/cranelift/filetests/filetests/isa/x64/fcvt-small.clif b/cranelift/filetests/filetests/isa/x64/fcvt-small.clif
new file mode 100644
index 0000000000..d49cea1d4d
--- /dev/null
+++ b/cranelift/filetests/filetests/isa/x64/fcvt-small.clif
@@ -0,0 +1,390 @@
+test compile precise-output
+set unwind_info=false
+target x86_64
+
+function %fcvt_to_uint_i8_f32(f32) -> i8 {
+block0(v0: f32):
+ v1 = fcvt_to_uint.i8 v0
+ return v1
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; cvt_float32_to_sint32_seq %xmm0, %eax, %rsi, %xmm3
+; cmpl $0xff, %eax
+; jnle #trap=int_ovf
+; cmpl $0x0, %eax
+; jl #trap=int_ovf
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; cvttss2si %xmm0, %eax
+; cmpl $1, %eax
+; jno 0x39
+; ucomiss %xmm0, %xmm0
+; jp 0x52
+; movl $0xcf000000, %esi
+; movd %esi, %xmm3
+; ucomiss %xmm3, %xmm0
+; jb 0x54
+; xorpd %xmm3, %xmm3
+; ucomiss %xmm0, %xmm3
+; jb 0x56
+; cmpl $0xff, %eax
+; jg 0x58
+; cmpl $0, %eax
+; jl 0x5a
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+; ud2 ; trap: bad_toint
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+
+function %fcvt_to_sint_i8_f32(f32) -> i8 {
+block0(v0: f32):
+ v1 = fcvt_to_sint.i8 v0
+ return v1
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; cvt_float32_to_sint32_seq %xmm0, %eax, %rsi, %xmm3
+; cmpl $0x7f, %eax
+; jnle #trap=int_ovf
+; cmpl $0xffffff80, %eax
+; jl #trap=int_ovf
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; cvttss2si %xmm0, %eax
+; cmpl $1, %eax
+; jno 0x39
+; ucomiss %xmm0, %xmm0
+; jp 0x50
+; movl $0xcf000000, %esi
+; movd %esi, %xmm3
+; ucomiss %xmm3, %xmm0
+; jb 0x52
+; xorpd %xmm3, %xmm3
+; ucomiss %xmm0, %xmm3
+; jb 0x54
+; cmpl $0x7f, %eax
+; jg 0x56
+; cmpl $-0x80, %eax
+; jl 0x58
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+; ud2 ; trap: bad_toint
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+
+function %fcvt_to_uint_sat_i8_f32(f32) -> i8 {
+block0(v0: f32):
+ v1 = fcvt_to_uint_sat.i8 v0
+ return v1
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; cvt_float32_to_sint32_sat_seq %xmm0, %eax, %rsi, %xmm3
+; movl $0xff, %esi
+; cmpl $0xff, %eax
+; cmovgl %esi, %eax
+; uninit %rcx
+; xorl %ecx, %ecx
+; cmpl $0x0, %eax
+; cmovll %ecx, %eax
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; cvttss2si %xmm0, %eax
+; cmpl $1, %eax
+; jno 0x33
+; ucomiss %xmm0, %xmm0
+; jnp 0x21
+; xorl %eax, %eax
+; jmp 0x33
+; xorpd %xmm3, %xmm3
+; ucomiss %xmm0, %xmm3
+; jae 0x33
+; movl $0x7fffffff, %eax
+; movl $0xff, %esi
+; cmpl $0xff, %eax
+; cmovgl %esi, %eax
+; xorl %ecx, %ecx
+; cmpl $0, %eax
+; cmovll %ecx, %eax
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+
+function %fcvt_to_sint_sat_i8_f32(f32) -> i8 {
+block0(v0: f32):
+ v1 = fcvt_to_sint_sat.i8 v0
+ return v1
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; cvt_float32_to_sint32_sat_seq %xmm0, %eax, %rsi, %xmm3
+; movl $0x7f, %esi
+; cmpl $0x7f, %eax
+; cmovgl %esi, %eax
+; movl $0xffffff80, %r11d
+; cmpl $0xffffff80, %eax
+; cmovll %r11d, %eax
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; cvttss2si %xmm0, %eax
+; cmpl $1, %eax
+; jno 0x33
+; ucomiss %xmm0, %xmm0
+; jnp 0x21
+; xorl %eax, %eax
+; jmp 0x33
+; xorpd %xmm3, %xmm3
+; ucomiss %xmm0, %xmm3
+; jae 0x33
+; movl $0x7fffffff, %eax
+; movl $0x7f, %esi
+; cmpl $0x7f, %eax
+; cmovgl %esi, %eax
+; movl $0xffffff80, %r11d
+; cmpl $-0x80, %eax
+; cmovll %r11d, %eax
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+
+function %fcvt_to_uint_i16_f64(f64) -> i16 {
+block0(v0: f64):
+ v1 = fcvt_to_uint.i16 v0
+ return v1
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; cvt_float64_to_sint32_seq %xmm0, %eax, %rsi, %xmm3
+; cmpl $0xffff, %eax
+; jnle #trap=int_ovf
+; cmpl $0x0, %eax
+; jl #trap=int_ovf
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; cvttsd2si %xmm0, %eax
+; cmpl $1, %eax
+; jno 0x42
+; ucomisd %xmm0, %xmm0
+; jp 0x5b
+; movabsq $13970166044105375744, %rsi
+; movq %rsi, %xmm3
+; ucomisd %xmm3, %xmm0
+; jbe 0x5d
+; xorpd %xmm3, %xmm3
+; ucomisd %xmm0, %xmm3
+; jb 0x5f
+; cmpl $0xffff, %eax
+; jg 0x61
+; cmpl $0, %eax
+; jl 0x63
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+; ud2 ; trap: bad_toint
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+
+function %fcvt_to_sint_i16_f64(f64) -> i16 {
+block0(v0: f64):
+ v1 = fcvt_to_sint.i16 v0
+ return v1
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; cvt_float64_to_sint32_seq %xmm0, %eax, %rsi, %xmm3
+; cmpl $0x7fff, %eax
+; jnle #trap=int_ovf
+; cmpl $0xffff8000, %eax
+; jl #trap=int_ovf
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; cvttsd2si %xmm0, %eax
+; cmpl $1, %eax
+; jno 0x42
+; ucomisd %xmm0, %xmm0
+; jp 0x5d
+; movabsq $13970166044105375744, %rsi
+; movq %rsi, %xmm3
+; ucomisd %xmm3, %xmm0
+; jbe 0x5f
+; xorpd %xmm3, %xmm3
+; ucomisd %xmm0, %xmm3
+; jb 0x61
+; cmpl $0x7fff, %eax
+; jg 0x63
+; cmpl $0xffff8000, %eax
+; jl 0x65
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+; ud2 ; trap: bad_toint
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+; ud2 ; trap: int_ovf
+
+function %fcvt_to_uint_sat_i16_f64(f64) -> i16 {
+block0(v0: f64):
+ v1 = fcvt_to_uint_sat.i16 v0
+ return v1
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; cvt_float64_to_sint32_sat_seq %xmm0, %eax, %rsi, %xmm3
+; movl $0xffff, %esi
+; cmpl $0xffff, %eax
+; cmovgl %esi, %eax
+; uninit %rcx
+; xorl %ecx, %ecx
+; cmpl $0x0, %eax
+; cmovll %ecx, %eax
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; cvttsd2si %xmm0, %eax
+; cmpl $1, %eax
+; jno 0x35
+; ucomisd %xmm0, %xmm0
+; jnp 0x22
+; xorl %eax, %eax
+; jmp 0x35
+; xorpd %xmm3, %xmm3
+; ucomisd %xmm0, %xmm3
+; jae 0x35
+; movl $0x7fffffff, %eax
+; movl $0xffff, %esi
+; cmpl $0xffff, %eax
+; cmovgl %esi, %eax
+; xorl %ecx, %ecx
+; cmpl $0, %eax
+; cmovll %ecx, %eax
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+
+function %fcvt_to_sint_sat_i16_f64(f64) -> i16 {
+block0(v0: f64):
+ v1 = fcvt_to_sint_sat.i16 v0
+ return v1
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; cvt_float64_to_sint32_sat_seq %xmm0, %eax, %rsi, %xmm3
+; movl $0x7fff, %esi
+; cmpl $0x7fff, %eax
+; cmovgl %esi, %eax
+; movl $0xffff8000, %r11d
+; cmpl $0xffff8000, %eax
+; cmovll %r11d, %eax
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; cvttsd2si %xmm0, %eax
+; cmpl $1, %eax
+; jno 0x35
+; ucomisd %xmm0, %xmm0
+; jnp 0x22
+; xorl %eax, %eax
+; jmp 0x35
+; xorpd %xmm3, %xmm3
+; ucomisd %xmm0, %xmm3
+; jae 0x35
+; movl $0x7fffffff, %eax
+; movl $0x7fff, %esi
+; cmpl $0x7fff, %eax
+; cmovgl %esi, %eax
+; movl $0xffff8000, %r11d
+; cmpl $0xffff8000, %eax
+; cmovll %r11d, %eax
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+
diff --git a/cranelift/filetests/filetests/runtests/conversion-i8-i16.clif b/cranelift/filetests/filetests/runtests/conversion-i8-i16.clif
index dde4635be5..a511ea6f8f 100644
--- a/cranelift/filetests/filetests/runtests/conversion-i8-i16.clif
+++ b/cranelift/filetests/filetests/runtests/conversion-i8-i16.clif
@@ -2,6 +2,8 @@ test interpret
test run
target aarch64
target s390x
+target x86_64
+target x86_64 has_avx
target riscv64 has_c has_zcb
target riscv64
@@ -17,6 +19,7 @@ block0(v0: f32):
; run: %f32_to_i8(-0x80.1) == 0x80
; run: %f32_to_i8(0x7f.1) == 0x7f
; run: %f32_to_i8(-0x0.1) == 0
+; run: %f32_to_i8(-0x80.f) == 0x80
function %f64_to_i8(f64) -> i8 {
block0(v0: f64):
@@ -56,6 +59,7 @@ block0(v0: f64):
; run: %f64_to_i16(-0x8000.1) == 0x8000
; run: %f64_to_i16(0x7fff.1) == 0x7fff
; run: %f64_to_i16(-0x0.1) == 0
+; run: %f64_to_i16(0x7fff.f) == 0x7fff
function %f32_to_u8(f32) -> i8 {
block0(v0:f32):
@@ -67,6 +71,7 @@ block0(v0:f32):
; run: %f32_to_u8(0x4.2) == 4
; run: %f32_to_u8(0x4.6) == 4
; run: %f32_to_u8(-0x0.1) == 0
+; run: %f32_to_u8(0xff.f) == 0xff
function %f64_to_u8(f64) -> i8 {
block0(v0:f64):
@@ -78,6 +83,7 @@ block0(v0:f64):
; run: %f64_to_u8(0x4.2) == 4
; run: %f64_to_u8(0x4.6) == 4
; run: %f64_to_u8(-0x0.1) == 0
+; run: %f64_to_u8(0xff.f) == 0xff
function %f32_to_u16(f32) -> i16 {
block0(v0:f32):
@@ -89,6 +95,7 @@ block0(v0:f32):
; run: %f32_to_u16(0x4.2) == 4
; run: %f32_to_u16(0x4.6) == 4
; run: %f32_to_u16(-0x0.1) == 0
+; run: %f32_to_u16(0xffff.f) == 0xffff
function %f64_to_u16(f64) -> i16 {
block0(v0:f64):
@@ -100,6 +107,7 @@ block0(v0:f64):
; run: %f64_to_u16(0x4.2) == 4
; run: %f64_to_u16(0x4.6) == 4
; run: %f64_to_u16(-0x0.1) == 0
+; run: %f64_to_u16(0xffff.f) == 0xffff
function %f32_to_i8_sat(f32) -> i8 {
block0(v0: f32):
diff --git a/cranelift/filetests/filetests/runtests/fcvt-sat-small.clif b/cranelift/filetests/filetests/runtests/fcvt-sat-small.clif
index 977336e903..9dffd9dcf7 100644
--- a/cranelift/filetests/filetests/runtests/fcvt-sat-small.clif
+++ b/cranelift/filetests/filetests/runtests/fcvt-sat-small.clif
@@ -1,7 +1,9 @@
+test interpret
test run
target aarch64
target s390x
-; x86_64 does not support `fcvt_to_{u,s}int_sat` to integers < 32 bits.
+target x86_64
+target x86_64 has_avx
function %fcvt_to_sint_sat_i8(f32) -> i8 {
block0(v0: f32):
@@ -96,7 +98,7 @@ block0(v0: f64):
; run: %fcvt_to_uint_sat_i8_f64(0xB2D05E00.0) == 255
; run: %fcvt_to_uint_sat_i8_f64(-0xB2D05E00.0) == 0
; run: %fcvt_to_uint_sat_i8_f64(0x1.fffffffffffffp1023) == 255
-; run: %fcvt_to_sint_sat_i8_f64(-0x1.fffffffffffffp1023) == 0
+; run: %fcvt_to_uint_sat_i8_f64(-0x1.fffffffffffffp1023) == 0
; run: %fcvt_to_uint_sat_i8_f64(NaN) == 0
function %fcvt_to_sint_sat_i16_f64(f64) -> i16 {
diff --git a/cranelift/filetests/filetests/runtests/issue-5992.clif b/cranelift/filetests/filetests/runtests/issue-5992.clif
index a1d9e36926..ac6b7e6e51 100644
--- a/cranelift/filetests/filetests/runtests/issue-5992.clif
+++ b/cranelift/filetests/filetests/runtests/issue-5992.clif
@@ -3,6 +3,7 @@ test run
target riscv64
target aarch64
target s390x
+target x86_64
function %a_f32(f32) -> i16 {
block0(v0: f32):
diff --git a/cranelift/filetests/filetests/runtests/issue-5993.clif b/cranelift/filetests/filetests/runtests/issue-5993.clif
index b19a865b24..ddd66dd87e 100644
--- a/cranelift/filetests/filetests/runtests/issue-5993.clif
+++ b/cranelift/filetests/filetests/runtests/issue-5993.clif
@@ -3,6 +3,7 @@ test run
target riscv64
target aarch64
target s390x
+target x86_64
function %a(f64) -> i8 {
block0(v0: f64):
diff --git a/cranelift/fuzzgen/src/function_generator.rs b/cranelift/fuzzgen/src/function_generator.rs
index 543ad4cc84..fd2a498206 100644
--- a/cranelift/fuzzgen/src/function_generator.rs
+++ b/cranelift/fuzzgen/src/function_generator.rs
@@ -497,7 +497,6 @@ fn valid_for_target(triple: &Triple, op: Opcode, args: &[Type], rets: &[Type]) -
(Opcode::Cls, &[I128], &[I128]),
// TODO
(Opcode::Bitselect, &[_, _, _], &[F32 | F64]),
- // https://github.com/bytecodealliance/wasmtime/issues/4897
// https://github.com/bytecodealliance/wasmtime/issues/4899
(
Opcode::FcvtToUint
@@ -505,7 +504,7 @@ fn valid_for_target(triple: &Triple, op: Opcode, args: &[Type], rets: &[Type]) -
| Opcode::FcvtToSint
| Opcode::FcvtToSintSat,
&[F32 | F64],
- &[I8 | I16 | I128]
+ &[I128]
),
(Opcode::FcvtToUint | Opcode::FcvtToSint, &[F32X4], &[I32X4]),
(