| Issue |
243395
|
| Summary |
[X86][AVX10.2] improve llvm.fptoui.sat and llvm.fptosi.sat (scalar+vector) avx10.2 codegen
|
| Labels |
good first issue,
backend:X86,
missed-optimization
|
| Assignees |
|
| Reporter |
Andarwinux
|
avx10.2 include native SAT F2I, but the coverage is not complete. It should be possible to use SAT F2I + SAT I2I to improve codegen.
https://godbolt.org/z/cE7qEeof1
```llvm
define i16 @test_unsigned_i16_f64(double %f) nounwind {
%x = call i16 @llvm.fptoui.sat(double %f)
ret i16 %x
}
define i16 @test_signed_i16_f64(double %f) nounwind {
%x = call i16 @llvm.fptosi.sat(double %f)
ret i16 %x
}
define i8 @test_unsigned_i8_f64(double %f) nounwind {
%x = call i8 @llvm.fptoui.sat(double %f)
ret i8 %x
}
define i8 @test_signed_i8_f64(double %f) nounwind {
%x = call i8 @llvm.fptosi.sat(double %f)
ret i8 %x
}
define i16 @test_unsigned_i16_f32(float %f) nounwind {
%x = call i16 @llvm.fptoui.sat(float %f)
ret i16 %x
}
define i16 @test_signed_i16_f32(float %f) nounwind {
%x = call i16 @llvm.fptosi.sat(float %f)
ret i16 %x
}
define i8 @test_unsigned_i8_f32(float %f) nounwind {
%x = call i8 @llvm.fptoui.sat(float %f)
ret i8 %x
}
define i8 @test_signed_i8_f32(float %f) nounwind {
%x = call i8 @llvm.fptosi.sat(float %f)
ret i8 %x
}
define i16 @test_unsigned_i16_f16(half %f) nounwind {
%x = call i16 @llvm.fptoui.sat(half %f)
ret i16 %x
}
define i16 @test_signed_i16_f16(half %f) nounwind {
%x = call i16 @llvm.fptosi.sat(half %f)
ret i16 %x
}
define i8 @test_unsigned_i8_f16(half %f) nounwind {
%x = call i8 @llvm.fptoui.sat(half %f)
ret i8 %x
}
define i8 @test_signed_i8_f16(half %f) nounwind {
%x = call i8 @llvm.fptosi.sat(half %f)
ret i8 %x
}
```
```asm
.LCPI0_0:
.quad 0x40efffe000000000 # double 65535
test_unsigned_i16_f64: # @test_unsigned_i16_f64
vxorpd xmm1, xmm1, xmm1
vmaxsd xmm0, xmm1, xmm0
vmovsd xmm1, qword ptr [rip + .LCPI0_0] # xmm1 = [6.5535E+4,0.0E+0]
vminsd xmm0, xmm1, xmm0
vcvttsd2si eax, xmm0
ret
.LCPI1_0:
.quad 0xc0e0000000000000 # double -32768
.LCPI1_1:
.quad 0x40dfffc000000000 # double 32767
test_signed_i16_f64: # @test_signed_i16_f64
vmovsd xmm1, qword ptr [rip + .LCPI1_0] # xmm1 = [-3.2768E+4,0.0E+0]
vmaxsd xmm0, xmm1, xmm0
vmovsd xmm1, qword ptr [rip + .LCPI1_1] # xmm1 = [3.2767E+4,0.0E+0]
vminsd xmm0, xmm1, xmm0
vcvttsd2si eax, xmm0
ret
.LCPI2_0:
.quad 0x406fe00000000000 # double 255
test_unsigned_i8_f64: # @test_unsigned_i8_f64
vxorpd xmm1, xmm1, xmm1
vmaxsd xmm0, xmm1, xmm0
vmovsd xmm1, qword ptr [rip + .LCPI2_0] # xmm1 = [2.55E+2,0.0E+0]
vminsd xmm0, xmm1, xmm0
vcvttsd2si eax, xmm0
ret
.LCPI3_0:
.quad 0xc060000000000000 # double -128
.LCPI3_1:
.quad 0x405fc00000000000 # double 127
test_signed_i8_f64: # @test_signed_i8_f64
vmovsd xmm1, qword ptr [rip + .LCPI3_0] # xmm1 = [-1.28E+2,0.0E+0]
vmaxsd xmm0, xmm1, xmm0
vmovsd xmm1, qword ptr [rip + .LCPI3_1] # xmm1 = [1.27E+2,0.0E+0]
vminsd xmm0, xmm1, xmm0
vcvttsd2si eax, xmm0
ret
.LCPI4_0:
.long 0x477fff00 # float 65535
test_unsigned_i16_f32: # @test_unsigned_i16_f32
vxorps xmm1, xmm1, xmm1
vmaxss xmm0, xmm1, xmm0
vmovss xmm1, dword ptr [rip + .LCPI4_0] # xmm1 = [6.5535E+4,0.0E+0,0.0E+0,0.0E+0]
vminss xmm0, xmm1, xmm0
vcvttss2si eax, xmm0
ret
.LCPI5_0:
.long 0xc7000000 # float -32768
.LCPI5_1:
.long 0x46fffe00 # float 32767
test_signed_i16_f32: # @test_signed_i16_f32
vmovss xmm1, dword ptr [rip + .LCPI5_0] # xmm1 = [-3.2768E+4,0.0E+0,0.0E+0,0.0E+0]
vmaxss xmm0, xmm1, xmm0
vmovss xmm1, dword ptr [rip + .LCPI5_1] # xmm1 = [3.2767E+4,0.0E+0,0.0E+0,0.0E+0]
vminss xmm0, xmm1, xmm0
vcvttss2si eax, xmm0
ret
.LCPI6_0:
.long 0x437f0000 # float 255
test_unsigned_i8_f32: # @test_unsigned_i8_f32
vxorps xmm1, xmm1, xmm1
vmaxss xmm0, xmm1, xmm0
vmovss xmm1, dword ptr [rip + .LCPI6_0] # xmm1 = [2.55E+2,0.0E+0,0.0E+0,0.0E+0]
vminss xmm0, xmm1, xmm0
vcvttss2si eax, xmm0
ret
.LCPI7_0:
.long 0xc3000000 # float -128
.LCPI7_1:
.long 0x42fe0000 # float 127
test_signed_i8_f32: # @test_signed_i8_f32
vmovss xmm1, dword ptr [rip + .LCPI7_0] # xmm1 = [-1.28E+2,0.0E+0,0.0E+0,0.0E+0]
vmaxss xmm0, xmm1, xmm0
vmovss xmm1, dword ptr [rip + .LCPI7_1] # xmm1 = [1.27E+2,0.0E+0,0.0E+0,0.0E+0]
vminss xmm0, xmm1, xmm0
vcvttss2si eax, xmm0
ret
.LCPI8_0:
.short 0x7bff # half 65504
test_unsigned_i16_f16: # @test_unsigned_i16_f16
vcvttsh2si eax, xmm0
xor ecx, ecx
vxorps xmm1, xmm1, xmm1
vucomish xmm0, xmm1
cmovb eax, ecx
vucomish xmm0, word ptr [rip + .LCPI8_0]
mov ecx, 65535
cmova eax, ecx
ret
.LCPI9_0:
.short 0xf800 # half -32768
.LCPI9_1:
.short 0x77ff # half 32752
test_signed_i16_f16: # @test_signed_i16_f16
vcvttsh2si eax, xmm0
vmovsh xmm1, word ptr [rip + .LCPI9_0] # xmm1 = [-3.2768E+4,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0]
vucomish xmm1, xmm0
mov ecx, 32768
cmova eax, ecx
vucomish xmm0, word ptr [rip + .LCPI9_1]
mov ecx, 32767
cmova eax, ecx
ret
.LCPI10_0:
.short 0x5bf8 # half 255
test_unsigned_i8_f16: # @test_unsigned_i8_f16
vxorps xmm1, xmm1, xmm1
vmaxsh xmm0, xmm1, xmm0
vmovsh xmm1, word ptr [rip + .LCPI10_0] # xmm1 = [2.55E+2,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0]
vminsh xmm0, xmm1, xmm0
vcvttsh2si eax, xmm0
ret
.LCPI11_0:
.short 0xd800 # half -128
.LCPI11_1:
.short 0x57f0 # half 127
test_signed_i8_f16: # @test_signed_i8_f16
vmovsh xmm1, word ptr [rip + .LCPI11_0] # xmm1 = [-1.28E+2,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0]
vmaxsh xmm0, xmm1, xmm0
vmovsh xmm1, word ptr [rip + .LCPI11_1] # xmm1 = [1.27E+2,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0,0.0E+0]
vminsh xmm0, xmm1, xmm0
vcvttsh2si eax, xmm0
ret
```
should be
```c
#include <stdint.h>
#include <immintrin.h>
uint16_t test_unsigned_i16_f64(__m128d a){
return _mm_cvtusepi32_epi16(_mm_cvtts_pd_epu32(a))[0];
}
int16_t test_signed_i16_f64(__m128d a){
return _mm_cvtsepi32_epi16(_mm_cvtts_pd_epi32(a))[0];
}
uint8_t test_unsigned_i8_f64(__m128d a){
return _mm_cvtusepi32_epi8(_mm_cvtts_pd_epu32(a))[0];
}
int8_t test_signed_i8_f64(__m128d a){
return _mm_cvtsepi32_epi8(_mm_cvtts_pd_epi32(a))[0];
}
uint16_t test_unsigned_i16_f32(__m128 a){
return _mm_cvtusepi32_epi16(_mm_cvtts_ps_epu32(a))[0];
}
int16_t test_signed_i16_f32(__m128 a){
return _mm_cvtsepi32_epi16(_mm_cvtts_ps_epi32(a))[0];
}
uint8_t test_unsigned_i8_f32(__m128 a){
return _mm_ipcvtts_ps_epu8(a)[0];
}
int8_t test_signed_i8_f32(__m128 a){
return _mm_ipcvtts_ps_epi8(a)[0];
}
uint16_t test_unsigned_i16_f16(__m128h a){
return _mm_cvtusepi32_epi16(_mm_cvtts_ps_epu32(_mm_cvtxph_ps(a)))[0];
}
int16_t test_signed_i16_f16(__m128h a){
return _mm_cvtsepi32_epi16(_mm_cvtts_ps_epi32(_mm_cvtxph_ps(a)))[0];
}
uint8_t test_unsigned_i8_f16(__m128h a){
return _mm_ipcvtts_ph_epu8(a)[0];
}
int8_t test_signed_i8_f16(__m128h a){
return _mm_ipcvtts_ph_epi8(a)[0];
}
```
```asm
test_unsigned_i16_f64:
vcvttpd2udqs xmm0, xmm0
vpmovusdw xmm0, xmm0
vmovw eax, xmm0
ret
test_signed_i16_f64:
vcvttpd2dqs xmm0, xmm0
vpmovsdw xmm0, xmm0
vmovw eax, xmm0
ret
test_unsigned_i8_f64:
vcvttpd2udqs xmm0, xmm0
vpmovusdb xmm0, xmm0
vmovd eax, xmm0
ret
test_signed_i8_f64:
vcvttpd2dqs xmm0, xmm0
vpmovsdb xmm0, xmm0
vmovd eax, xmm0
ret
test_unsigned_i16_f32:
vcvttps2udqs xmm0, xmm0
vpmovusdw xmm0, xmm0
vmovw eax, xmm0
ret
test_signed_i16_f32:
vcvttps2dqs xmm0, xmm0
vpmovsdw xmm0, xmm0
vmovw eax, xmm0
ret
test_unsigned_i8_f32:
vcvttps2iubs xmm0, xmm0
vmovd eax, xmm0
ret
test_signed_i8_f32:
vcvttps2ibs xmm0, xmm0
vmovd eax, xmm0
ret
test_unsigned_i16_f16:
vcvtph2psx xmm0, xmm0
vcvttps2udqs xmm0, xmm0
vpmovusdw xmm0, xmm0
vmovw eax, xmm0
ret
test_signed_i16_f16:
vcvtph2psx xmm0, xmm0
vcvttps2dqs xmm0, xmm0
vpmovsdw xmm0, xmm0
vmovw eax, xmm0
ret
test_unsigned_i8_f16:
vcvttph2iubs xmm0, xmm0
vmovd eax, xmm0
ret
test_signed_i8_f16:
vcvttph2ibs xmm0, xmm0
vmovd eax, xmm0
ret
```
This also works for vector, the only thing to note is that vcvttps2iubs/vcvttps2ibs/vcvttph2iubs/vcvttph2ibs output i32/i16, so still need to be truncated.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs