zstd: Shave some instructions off the amd64 asm (#720)

These are all in return sequences, so no performance gain expected, but the generated code is smaller.
klauspost · Dec 29, 2022 · 5206968 · 5206968
1 parent 272fbc7
commit 5206968
Show file tree

Hide file tree

Showing 2 changed files with 12 additions and 33 deletions.
diff --git a/zstd/_generate/gen.go b/zstd/_generate/gen.go
@@ -502,20 +502,20 @@ func (o options) generateBody(name string, executeSingleTriple func(ctx *execute
 		o.returnWithCode(errorNotEnoughLiterals)
 	}
 
-	Comment("Return with not enough output space error")
-	{
+	if !o.useSeqs {
+		Comment("Return with not enough output space error")
 		Label("error_not_enough_space")
-		if !o.useSeqs {
+		{
 			ctx := Dereference(Param("ctx"))
 			tmp := GP64()
 			MOVQ(llP, tmp)
 			Store(tmp, ctx.Field("ll"))
 			MOVQ(mlP, tmp)
 			Store(tmp, ctx.Field("ml"))
 			Store(ec.outPosition, ctx.Field("outPosition"))
-		}
 
-		o.returnWithCode(errorNotEnoughSpace)
+			o.returnWithCode(errorNotEnoughSpace)
+		}
 	}
 }
 
@@ -1062,10 +1062,9 @@ func (e executeSimple) generateProcedure(name string) {
 		Store(seqIndex, ctx.Field("seqIndex"))
 		Store(outPosition, ctx.Field("outPosition"))
 
-		// compute litPosition
-		tmp := GP64()
-		Load(ctx.Field("literals").Base(), tmp)
-		SUBQ(tmp, literals) // litPosition := current - initial literals pointer
+		// litPosition := current - initial literals pointer
+		litField, _ := ctx.Field("literals").Base().Resolve()
+		SUBQ(litField.Addr, literals)
 		Store(literals, ctx.Field("litPosition"))
 	}
 	Label("loop_finished")

diff --git a/zstd/seqdec_amd64.s b/zstd/seqdec_amd64.s
@@ -320,10 +320,6 @@ error_not_enough_literals:
 	MOVQ $0x00000004, ret+24(FP)
 	RET
 
-	// Return with not enough output space error
-	MOVQ $0x00000005, ret+24(FP)
-	RET
-
 // func sequenceDecs_decode_56_amd64(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
 // Requires: CMOV
 TEXT ·sequenceDecs_decode_56_amd64(SB), $8-32
@@ -617,10 +613,6 @@ error_not_enough_literals:
 	MOVQ $0x00000004, ret+24(FP)
 	RET
 
-	// Return with not enough output space error
-	MOVQ $0x00000005, ret+24(FP)
-	RET
-
 // func sequenceDecs_decode_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
 // Requires: BMI, BMI2, CMOV
 TEXT ·sequenceDecs_decode_bmi2(SB), $8-32
@@ -897,10 +889,6 @@ error_not_enough_literals:
 	MOVQ $0x00000004, ret+24(FP)
 	RET
 
-	// Return with not enough output space error
-	MOVQ $0x00000005, ret+24(FP)
-	RET
-
 // func sequenceDecs_decode_56_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
 // Requires: BMI, BMI2, CMOV
 TEXT ·sequenceDecs_decode_56_bmi2(SB), $8-32
@@ -1152,10 +1140,6 @@ error_not_enough_literals:
 	MOVQ $0x00000004, ret+24(FP)
 	RET
 
-	// Return with not enough output space error
-	MOVQ $0x00000005, ret+24(FP)
-	RET
-
 // func sequenceDecs_executeSimple_amd64(ctx *executeAsmContext) bool
 // Requires: SSE
 TEXT ·sequenceDecs_executeSimple_amd64(SB), $8-9
@@ -1389,8 +1373,7 @@ loop_finished:
 	MOVQ ctx+0(FP), AX
 	MOVQ DX, 24(AX)
 	MOVQ DI, 104(AX)
-	MOVQ 80(AX), CX
-	SUBQ CX, SI
+	SUBQ 80(AX), SI
 	MOVQ SI, 112(AX)
 	RET
 
@@ -1402,8 +1385,7 @@ error_match_off_too_big:
 	MOVQ ctx+0(FP), AX
 	MOVQ DX, 24(AX)
 	MOVQ DI, 104(AX)
-	MOVQ 80(AX), CX
-	SUBQ CX, SI
+	SUBQ 80(AX), SI
 	MOVQ SI, 112(AX)
 	RET
 
@@ -1747,8 +1729,7 @@ loop_finished:
 	MOVQ ctx+0(FP), AX
 	MOVQ DX, 24(AX)
 	MOVQ DI, 104(AX)
-	MOVQ 80(AX), CX
-	SUBQ CX, SI
+	SUBQ 80(AX), SI
 	MOVQ SI, 112(AX)
 	RET
 
@@ -1760,8 +1741,7 @@ error_match_off_too_big:
 	MOVQ ctx+0(FP), AX
 	MOVQ DX, 24(AX)
 	MOVQ DI, 104(AX)
-	MOVQ 80(AX), CX
-	SUBQ CX, SI
+	SUBQ 80(AX), SI
 	MOVQ SI, 112(AX)
 	RET