petrbalvin/gasm-sdk
0
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 37 Wiki Activity

feat(arch): scale and exponent-extract the scalar FP16 in the extension layer

Browse Source 
Assisted-by: GLM 5.3 Flash
This commit is contained in:
petrbalvin committed 2026-10-07 00:07:58 +02:00
1 parent 127e52de69
commit 0354a1f4c1
3 files changed
+20 -4
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

arch/amd64_ext.go
+6
View File
@@ -364,6 +364,12 @@ var amd64Extensions = []ExtInstr{
{Name: "VSQRTSH", Summary: "Compute the square root of a scalar FP16 value",
Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x51, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VSQRTSH (EVEX.NDS.LIG.F3.MAP5.W0 51 /r)"},
{Name: "VSCALEFSH", Summary: "Scale a scalar FP16 value by the ratio of two others",
Bytes: []byte{0x62, 0x06, 0x05, 0x00, 0x2D, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VSCALEFSH (EVEX.NDS.LIG.66.MAP6.W0 2D /r)"},
{Name: "VGETEXPSH", Summary: "Convert the exponent of a scalar FP16 value to an FP16 value",
Bytes: []byte{0x62, 0x06, 0x05, 0x00, 0x43, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VGETEXPSH (EVEX.NDS.LIG.66.MAP6.W0 43 /r)"},
{Name: "VCOMISH", Summary: "Compare a scalar FP16 value and set EFLAGS",
Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x2F, 0xC0}, Form: ExtFormAmdVec2, Feature: ExtFeatureFP16,
Ref: "Intel SDM Vol. 2C, VCOMISH (EVEX.LIG.MAP5.W0 2F /r)"},
arch/amd64_ext_test.go
+10 -2
View File
@@ -108,6 +108,14 @@ var amd64GoldenRows = []amd64GoldenRow{
[]ExtOperand{ExtXmm(29), ExtXmm(28), ExtXmm(30)},
"6205160051f4", "62 05 16 00 51 f4 vsqrtsh %xmm28,%xmm29,%xmm30"},
// The scalar scale and exponent extracts, EVEX.NDS.LIG.66.MAP6.W0.
{"vscalefsh", "VSCALEFSH",
[]ExtOperand{ExtXmm(29), ExtXmm(28), ExtXmm(30)},
"620615002df4", "62 06 15 00 2d f4 vscalefsh %xmm28,%xmm29,%xmm30"},
{"vgetexpsh", "VGETEXPSH",
[]ExtOperand{ExtXmm(29), ExtXmm(28), ExtXmm(30)},
"6206150043f4", "62 06 15 00 43 f4 vgetexpsh %xmm28,%xmm29,%xmm30"},
// The scalar compares take two operands, EVEX.LIG.MAP5.W0.
{"vcomish", "VCOMISH",
[]ExtOperand{ExtXmm(29), ExtXmm(30)},
@@ -416,7 +424,7 @@ func TestAmd64ExtArchBinding(t *testing.T) {
t.Errorf("Extensions(%s) carries %d instructions, want none", a, len(got))
}
}
if got := Extensions(AMD64); len(got) != 46 {
t.Errorf("the amd64 layer registers %d instructions, want 46", len(got))
if got := Extensions(AMD64); len(got) != 48 {
t.Errorf("the amd64 layer registers %d instructions, want 48", len(got))
}
}
asm/extension_amd64_test.go
+4 -2
View File
@@ -35,6 +35,8 @@ func TestAmd64ExtensionRegistry(t *testing.T) {
{"VCVTSH2SI", 2},
{"VADDPH", 1},
{"VSQRTPH", 1},
{"VSCALEFSH", 1},
{"VGETEXPSH", 1},
} {
cands, ok := LookupExtension(arch.AMD64, tt.mnem)
if !ok {
@@ -48,8 +50,8 @@ func TestAmd64ExtensionRegistry(t *testing.T) {
t.Errorf("the %s lookup is not case-insensitive", tt.mnem)
}
}
if got := arch.Extensions(arch.AMD64); len(got) != 46 {
t.Errorf("the amd64 layer registers %d instructions, want 46", len(got))
if got := arch.Extensions(arch.AMD64); len(got) != 48 {
t.Errorf("the amd64 layer registers %d instructions, want 48", len(got))
}
if _, ok := LookupExtension(arch.AMD64, "NOSUCHINSTR"); ok {
t.Error("a non-extended mnemonic resolved")